diff --git a/Cargo.toml b/Cargo.toml
index da79e13..e47ec79 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -54,6 +54,7 @@ cust_raw = { git = "https://github.com/Rust-GPU/Rust-CUDA", branch = "main", def
 cuda-runtime-sys = { version = "0.3.0-alpha.1", optional = true}
 cudarc = { version = "0.18.2", features = ["cuda-13000"], optional = true }
 jemallocator = { version = "0.5", optional = true }
+syn = "2.0.117"
 
 [build-dependencies]
 spirv-builder = { git = "https://github.com/rust-gpu/rust-gpu", branch = "main", optional = true }
diff --git a/crates/ptex-filter/Cargo.toml b/crates/ptex-filter/Cargo.toml
deleted file mode 100644
index f0a52e7..0000000
--- a/crates/ptex-filter/Cargo.toml
+++ /dev/null
@@ -1,13 +0,0 @@
-#![allow(unused)]
-
-[package]
-name = "ptex-filter"
-version = "0.1.0"
-edition = "2021"
-
-[build-dependencies]
-cc = "1.0"
-pkg-config = "0.3"
-
-[dependencies]
-ptex = "0.3.0"
diff --git a/crates/ptex-filter/build.rs b/crates/ptex-filter/build.rs
deleted file mode 100644
index 0b6774c..0000000
--- a/crates/ptex-filter/build.rs
+++ /dev/null
@@ -1,17 +0,0 @@
-fn main() {
-    println!("cargo:rerun-if-changed=cpp/ptex_filter_wrapper.cpp");
-
-    let home = std::env::var("HOME").unwrap();
-    let ptex_include = format!("{}/.local/include", home);
-    let ptex_lib = format!("{}/.local/lib", home);
-
-    cc::Build::new()
-        .cpp(true)
-        .file("cpp/ptex_filter_wrapper.cpp")
-        .include(&ptex_include)
-        .flag("-std=c++17")
-        .compile("ptex_filter_wrapper");
-
-    println!("cargo:rustc-link-search=native={}", ptex_lib);
-    println!("cargo:rustc-link-lib=Ptex");
-}
diff --git a/crates/ptex-filter/cpp/ptex_filter_wrapper.cpp b/crates/ptex-filter/cpp/ptex_filter_wrapper.cpp
deleted file mode 100644
index b2a119c..0000000
--- a/crates/ptex-filter/cpp/ptex_filter_wrapper.cpp
+++ /dev/null
@@ -1,65 +0,0 @@
-#include "ptex_filter_wrapper.h"
-#include <Ptexture.h>
-
-extern "C" {
-
-PtexFilterHandle ptex_filter_create(PtexTextureHandle texture, const PtexFilterOptions* opts) {
-    Ptex::PtexTexture* tex = static_cast<Ptex::PtexTexture*>(texture);
-    if (!tex || !opts) return nullptr;
-
-    Ptex::PtexFilter::Options ptex_opts;
-    ptex_opts.filter = static_cast<Ptex::PtexFilter::FilterType>(opts->filter);
-    ptex_opts.lerp = opts->lerp;
-    ptex_opts.sharpness = opts->sharpness;
-    ptex_opts.noedgeblend = opts->noedgeblend != 0;
-
-    return Ptex::PtexFilter::getFilter(tex, ptex_opts);
-}
-
-void ptex_filter_eval(
-    PtexFilterHandle filter,
-    float* result,
-    int32_t first_channel,
-    int32_t num_channels,
-    int32_t face_id,
-    float u, float v,
-    float dudx, float dvdx,
-    float dudy, float dvdy
-) {
-    Ptex::PtexFilter* f = static_cast<Ptex::PtexFilter*>(filter);
-    if (f && result) {
-        f->eval(result, first_channel, num_channels, face_id, u, v, dudx, dvdx, dudy, dvdy);
-    }
-}
-
-void ptex_filter_release(PtexFilterHandle filter) {
-    Ptex::PtexFilter* f = static_cast<Ptex::PtexFilter*>(filter);
-    if (f) {
-        f->release();
-    }
-}
-
-PtexTextureHandle ptex_texture_open(const char* filename, char** error_str) {
-    Ptex::String error;
-    Ptex::PtexTexture* tex = Ptex::PtexTexture::open(filename, error);
-    
-    if (!tex && error_str) {
-        *error_str = strdup(error.c_str());
-    }
-    
-    return tex;
-}
-
-void ptex_texture_release(PtexTextureHandle texture) {
-    Ptex::PtexTexture* tex = static_cast<Ptex::PtexTexture*>(texture);
-    if (tex) {
-        tex->release();
-    }
-}
-
-int32_t ptex_texture_num_channels(PtexTextureHandle texture) {
-    Ptex::PtexTexture* tex = static_cast<Ptex::PtexTexture*>(texture);
-    return tex ? tex->numChannels() : 0;
-}
-
-}
diff --git a/crates/ptex-filter/cpp/ptex_filter_wrapper.h b/crates/ptex-filter/cpp/ptex_filter_wrapper.h
deleted file mode 100644
index b999325..0000000
--- a/crates/ptex-filter/cpp/ptex_filter_wrapper.h
+++ /dev/null
@@ -1,51 +0,0 @@
-#pragma once
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#include <stdint.h>
-
-typedef void* PtexTextureHandle;
-typedef void* PtexFilterHandle;
-
-typedef enum {
-    PTEX_FILTER_POINT = 0,
-    PTEX_FILTER_BILINEAR = 1,
-    PTEX_FILTER_BOX = 2,
-    PTEX_FILTER_GAUSSIAN = 3,
-    PTEX_FILTER_BICUBIC = 4,
-    PTEX_FILTER_BSPLINE = 5,
-    PTEX_FILTER_CATMULLROM = 6,
-    PTEX_FILTER_MITCHELL = 7
-} PtexFilterType;
-
-typedef struct {
-    PtexFilterType filter;
-    int32_t lerp;
-    float sharpness;
-    int32_t noedgeblend;
-} PtexFilterOptions;
-
-PtexTextureHandle ptex_texture_open(const char* filename, char** error_str);
-void ptex_filter_release(PtexFilterHandle filter);
-int32_t ptex_texture_num_channels(PtexTextureHandle texture);
-PtexFilterHandle ptex_filter_create(PtexTextureHandle texture, const PtexFilterOptions* opts);
-
-
-void ptex_filter_eval(
-    PtexFilterHandle filter,
-    float* result,
-    int32_t first_channel,
-    int32_t num_channels,
-    int32_t face_id,
-    float u, float v,
-    float dudx, float dvdx,
-    float dudy, float dvdy
-);
-void ptex_filter_release(PtexFilterHandle filter);
-
-
-#ifdef __cplusplus
-}
-#endif
diff --git a/crates/ptex-filter/src/ffi.rs b/crates/ptex-filter/src/ffi.rs
deleted file mode 100644
index be7b391..0000000
--- a/crates/ptex-filter/src/ffi.rs
+++ /dev/null
@@ -1,60 +0,0 @@
-use std::ffi::c_void;
-
-#[repr(C)]
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum PtexFilterType {
-    Point = 0,
-    Bilinear = 1,
-    Box = 2,
-    Gaussian = 3,
-    Bicubic = 4,
-    BSpline = 5,
-    CatmullRom = 6,
-    Mitchell = 7,
-}
-
-#[repr(C)]
-#[derive(Debug, Clone)]
-pub struct PtexFilterOptions {
-    pub filter: PtexFilterType,
-    pub lerp: i32,
-    pub sharpness: f32,
-    pub noedgeblend: i32,
-}
-
-impl Default for PtexFilterOptions {
-    fn default() -> Self {
-        Self {
-            filter: PtexFilterType::BSpline,
-            lerp: 1,
-            sharpness: 0.0,
-            noedgeblend: 0,
-        }
-    }
-}
-
-extern "C" {
-    pub fn ptex_filter_create(texture: *mut c_void, opts: *const PtexFilterOptions) -> *mut c_void;
-
-    pub fn ptex_filter_eval(
-        filter: *mut c_void,
-        result: *mut f32,
-        first_channel: i32,
-        num_channels: i32,
-        face_id: i32,
-        u: f32,
-        v: f32,
-        dudx: f32,
-        dvdx: f32,
-        dudy: f32,
-        dvdy: f32,
-    );
-
-    pub fn ptex_filter_release(filter: *mut c_void);
-    pub fn ptex_texture_open(
-        filename: *const std::ffi::c_char,
-        error_str: *mut *mut std::ffi::c_char,
-    ) -> *mut c_void;
-    pub fn ptex_texture_release(texture: *mut c_void);
-    pub fn ptex_texture_num_channels(texture: *mut c_void) -> i32;
-}
diff --git a/crates/ptex-filter/src/lib.rs b/crates/ptex-filter/src/lib.rs
deleted file mode 100644
index 38dbc27..0000000
--- a/crates/ptex-filter/src/lib.rs
+++ /dev/null
@@ -1,66 +0,0 @@
-#![allow(unused)]
-#![allow(dead_code)]
-
-mod ffi;
-
-pub use ffi::{ptex_filter_create, PtexFilterOptions, PtexFilterType};
-
-use std::ffi::c_void;
-use std::marker::PhantomData;
-use std::ptr::NonNull;
-
-pub struct PtexFilter {
-    handle: NonNull<c_void>,
-    _marker: PhantomData<*mut ()>,
-}
-
-impl PtexFilter {
-    /// Creates a new Ptex filter pointer
-    ///
-    /// # Safety
-    pub unsafe fn new(texture_ptr: *mut c_void, opts: &PtexFilterOptions) -> Option<Self> {
-        let handle = ffi::ptex_filter_create(texture_ptr, opts);
-        NonNull::new(handle).map(|h| Self {
-            handle: h,
-            _marker: PhantomData,
-        })
-    }
-
-    pub fn eval(
-        &self,
-        face_id: i32,
-        u: f32,
-        v: f32,
-        dudx: f32,
-        dvdx: f32,
-        dudy: f32,
-        dvdy: f32,
-        num_channels: i32,
-    ) -> [f32; 4] {
-        let mut result = [0.0f32; 4];
-        unsafe {
-            ffi::ptex_filter_eval(
-                self.handle.as_ptr(),
-                result.as_mut_ptr(),
-                0,
-                num_channels.min(4),
-                face_id,
-                u,
-                v,
-                dudx,
-                dvdx,
-                dudy,
-                dvdy,
-            );
-        }
-        result
-    }
-}
-
-impl Drop for PtexFilter {
-    fn drop(&mut self) {
-        unsafe {
-            ffi::ptex_filter_release(self.handle.as_ptr());
-        }
-    }
-}
diff --git a/shared/src/cameras/realistic.rs b/shared/src/cameras/realistic.rs
index 96440ce..0f513da 100644
--- a/shared/src/cameras/realistic.rs
+++ b/shared/src/cameras/realistic.rs
@@ -1,18 +1,16 @@
-use crate::PI;
 use crate::core::camera::{CameraBase, CameraRay, CameraTrait, CameraTransform};
 use crate::core::color::SRGB;
 use crate::core::film::Film;
 use crate::core::geometry::{
     Bounds2f, Normal3f, Point2f, Point2i, Point3f, Ray, Vector2f, Vector2i, Vector3f, VectorLike,
 };
-use crate::core::image::{DeviceImage, PixelFormat};
+use crate::core::image::{Image, PixelFormat};
 use crate::core::medium::Medium;
-use crate::core::pbrt::Float;
 use crate::core::sampler::CameraSample;
 use crate::core::scattering::refract;
 use crate::spectra::{SampledSpectrum, SampledWavelengths};
-use crate::utils::Ptr;
 use crate::utils::math::{lerp, quadratic, square};
+use crate::{Float, GVec, Ptr, PI};
 use num_traits::Float as NumFloat;
 
 #[repr(C)]
@@ -39,8 +37,8 @@ pub struct RealisticCamera {
     pub base: CameraBase,
     pub focus_distance: Float,
     pub set_aperture_diameter: Float,
-    pub aperture_image: Ptr<DeviceImage>,
-    pub element_interfaces: Ptr<LensElementInterface>,
+    pub aperture_image: Ptr<Image>,
+    pub element_interfaces: GVec<LensElementInterface>,
     pub n_elements: usize,
     pub physical_extent: Bounds2f,
     pub exit_pupil_bounds: [Bounds2f; EXIT_PUPIL_SAMPLES],
diff --git a/shared/src/core/aggregates.rs b/shared/src/core/aggregates.rs
index 94f3154..885b00d 100644
--- a/shared/src/core/aggregates.rs
+++ b/shared/src/core/aggregates.rs
@@ -2,6 +2,7 @@ use crate::core::geometry::{Bounds3f, Ray, Vector3f};
 use crate::core::pbrt::Float;
 use crate::core::primitive::{Primitive, PrimitiveTrait};
 use crate::core::shape::ShapeIntersection;
+use crate::{Float, Ptr, GVec, gvec};
 use crate::utils::Ptr;
 
 #[repr(C)]
@@ -15,40 +16,40 @@ pub struct LinearBVHNode {
 }
 
 #[repr(C)]
-#[derive(Debug, Clone, Copy)]
-pub struct DeviceBVHAggregate {
+#[derive(Debug, Clone)]
+pub struct BVHAggregate {
     pub max_prims_in_node: u32,
-    pub primitives: Ptr<Primitive>,
+    pub primitives: GVec<Primitive>,
     pub primitive_count: u32,
-    pub nodes: Ptr<LinearBVHNode>,
+    pub nodes: GVec<LinearBVHNode>,
     pub node_count: u32,
 }
 
-impl DeviceBVHAggregate {
-    pub const fn empty() -> Self {
+impl BVHAggregate {
+    pub fn empty() -> Self {
         Self {
             max_prims_in_node: 0,
-            primitives: Ptr::null(),
+            primitives: gvec(),
             primitive_count: 0,
-            nodes: Ptr::null(),
+            nodes: gvec(),
             node_count: 0,
         }
     }
 
     #[inline(always)]
     fn node(&self, i: usize) -> &LinearBVHNode {
-        unsafe { self.nodes.at(i) }
+        unsafe { self.nodes.get_unchecked(i) }
     }
 
     #[inline(always)]
     fn primitive(&self, i: usize) -> &Primitive {
-        unsafe { self.primitives.at(i) }
+        unsafe { self.primitives.get_unchecked(i) }
     }
 }
 
-impl PrimitiveTrait for DeviceBVHAggregate {
+impl PrimitiveTrait for BVHAggregate {
     fn bounds(&self) -> Bounds3f {
-        if self.nodes.is_null() || self.node_count == 0 {
+        if self.nodes.is_empty() || self.node_count == 0 {
             Bounds3f::default()
         } else {
             self.node(0).bounds
@@ -56,7 +57,7 @@ impl PrimitiveTrait for DeviceBVHAggregate {
     }
 
     fn intersect(&self, r: &Ray, t_max: Option<Float>) -> Option<ShapeIntersection> {
-        if self.nodes.is_null() {
+        if self.nodes.is_empty() {
             return None;
         }
 
@@ -124,7 +125,7 @@ impl PrimitiveTrait for DeviceBVHAggregate {
     }
 
     fn intersect_p(&self, r: &Ray, t_max: Option<Float>) -> bool {
-        if self.nodes.is_null() || self.node_count == 0 {
+        if self.nodes.is_empty() || self.node_count == 0 {
             return false;
         }
 
diff --git a/shared/src/core/film.rs b/shared/src/core/film.rs
index 1952437..906ab66 100644
--- a/shared/src/core/film.rs
+++ b/shared/src/core/film.rs
@@ -1,24 +1,23 @@
 use crate::core::camera::CameraTransform;
-use crate::core::color::{MatrixMulColor, RGB, SRGB, XYZ, white_balance};
+use crate::core::color::{white_balance, MatrixMulColor, RGB, SRGB, XYZ};
 use crate::core::filter::{Filter, FilterTrait};
 use crate::core::geometry::{
     Bounds2f, Bounds2fi, Bounds2i, Normal3f, Point2f, Point2i, Point3f, Tuple, Vector2f, Vector2fi,
     Vector2i, Vector3f,
 };
-use crate::core::image::{DeviceImage, PixelFormat};
+use crate::core::image::{Image, PixelFormat};
 use crate::core::interaction::SurfaceInteraction;
-use crate::core::pbrt::Float;
 use crate::core::spectrum::{Spectrum, SpectrumTrait, StandardSpectra};
 use crate::spectra::{
-    ConstantSpectrum, DenselySampledSpectrum, LAMBDA_MAX, LAMBDA_MIN, N_SPECTRUM_SAMPLES,
-    PiecewiseLinearSpectrum, RGBColorSpace, SampledSpectrum, SampledWavelengths, colorspace,
+    colorspace, ConstantSpectrum, DenselySampledSpectrum, PiecewiseLinearSpectrum, RGBColorSpace,
+    SampledSpectrum, SampledWavelengths, LAMBDA_MAX, LAMBDA_MIN, N_SPECTRUM_SAMPLES,
 };
-use crate::utils::containers::DeviceArray2D;
 use crate::utils::math::linear_least_squares;
-use crate::utils::math::{SquareMatrix, wrap_equal_area_square};
+use crate::utils::math::{wrap_equal_area_square, SquareMatrix};
 use crate::utils::sampling::VarianceEstimator;
 use crate::utils::transform::AnimatedTransform;
-use crate::utils::{AtomicFloat, Ptr, gpu_array_from_fn};
+use crate::utils::{gpu_array_from_fn, AtomicFloat};
+use crate::{Array2D, Float, Ptr};
 use num_traits::Float as NumFloat;
 
 #[repr(C)]
@@ -29,7 +28,7 @@ pub struct RGBFilm {
     pub write_fp16: bool,
     pub filter_integral: Float,
     pub output_rgbf_from_sensor_rgb: SquareMatrix<Float, 3>,
-    pub pixels: DeviceArray2D<RGBPixel>,
+    pub pixels: Array2D<RGBPixel>,
 }
 
 #[repr(C)]
@@ -59,7 +58,7 @@ impl RGBFilm {
         &mut self.base
     }
 
-    pub fn get_sensor(&self) -> &DevicePixelSensor {
+    pub fn get_sensor(&self) -> &PixelSensor {
         #[cfg(not(target_os = "cuda"))]
         {
             if self.base.sensor.is_null() {
@@ -206,7 +205,7 @@ pub struct GBufferFilm {
     pub base: FilmBase,
     pub output_from_render: AnimatedTransform,
     pub apply_inverse: bool,
-    pub pixels: DeviceArray2D<GBufferPixel>,
+    pub pixels: Array2D<GBufferPixel>,
     pub colorspace: RGBColorSpace,
     pub max_component_value: Float,
     pub write_fp16: bool,
@@ -223,7 +222,7 @@ impl GBufferFilm {
         &mut self.base
     }
 
-    pub fn get_sensor(&self) -> &DevicePixelSensor {
+    pub fn get_sensor(&self) -> &PixelSensor {
         #[cfg(not(target_os = "cuda"))]
         {
             if self.base.sensor.is_null() {
@@ -359,7 +358,7 @@ pub struct SpectralFilm {
     pub max_component_value: Float,
     pub write_fp16: bool,
     pub filter_integral: Float,
-    pub pixels: DeviceArray2D<SpectralPixel>,
+    pub pixels: Array2D<SpectralPixel>,
     pub output_rgbf_from_sensor_rgb: SquareMatrix<Float, 3>,
     pub bucket_sums: *mut f64,
     pub weight_sums: *mut f64,
@@ -404,7 +403,7 @@ impl SpectralFilm {
 
 #[repr(C)]
 #[derive(Debug, Copy, Clone)]
-pub struct DevicePixelSensor {
+pub struct PixelSensor {
     pub xyz_from_sensor_rgb: SquareMatrix<Float, 3>,
     pub r_bar: DenselySampledSpectrum,
     pub g_bar: DenselySampledSpectrum,
@@ -412,7 +411,7 @@ pub struct DevicePixelSensor {
     pub imaging_ratio: Float,
 }
 
-impl DevicePixelSensor {
+impl PixelSensor {
     pub fn project_reflectance<T>(
         refl: &Spectrum,
         illum: &Spectrum,
@@ -492,7 +491,7 @@ pub struct FilmBase {
     pub pixel_bounds: Bounds2i,
     pub filter: Filter,
     pub diagonal: Float,
-    pub sensor: Ptr<DevicePixelSensor>,
+    pub sensor: Ptr<PixelSensor>,
 }
 
 #[repr(C)]
diff --git a/shared/src/core/filter.rs b/shared/src/core/filter.rs
index 740c036..b783e6a 100644
--- a/shared/src/core/filter.rs
+++ b/shared/src/core/filter.rs
@@ -1,9 +1,8 @@
 use crate::core::geometry::{Bounds2f, Bounds2i, Point2f, Point2i, Vector2f};
-use crate::core::pbrt::Float;
 use crate::filters::*;
-use crate::utils::containers::DeviceArray2D;
 use crate::utils::math::{gaussian, gaussian_integral, lerp, sample_tent, windowed_sinc};
-use crate::utils::sampling::DevicePiecewiseConstant2D;
+use crate::utils::sampling::PiecewiseConstant2D;
+use crate::{DeviceArray2D, Float};
 use enum_dispatch::enum_dispatch;
 
 pub struct FilterSample {
@@ -13,22 +12,44 @@ pub struct FilterSample {
 
 #[repr(C)]
 #[derive(Clone, Debug, Copy)]
-pub struct DeviceFilterSampler {
+pub struct FilterSampler {
     pub domain: Bounds2f,
-    pub distrib: DevicePiecewiseConstant2D,
-    pub f: DeviceArray2D<Float>,
+    pub distrib: PiecewiseConstant2D,
+    pub f: Array2D<Float>,
 }
 
-impl DeviceFilterSampler {
+impl FilterSampler {
+    pub fn new<F>(radius: Vector2f, func: F) -> Self
+    where
+        F: Fn(Point2f) -> Float,
+    {
+        let domain = Bounds2f::from_points(
+            Point2f::new(-radius.x(), -radius.y()),
+            Point2f::new(radius.x(), radius.y()),
+        );
+        let nx = (32.0 * radius.x()) as i32;
+        let ny = (32.0 * radius.y()) as i32;
+        let mut f = Array2D::new_dims(nx, ny);
+        for y in 0..f.y_size() {
+            for x in 0..f.x_size() {
+                let p = domain.lerp(Point2f::new(
+                    (x as Float + 0.5) / f.x_size() as Float,
+                    (y as Float + 0.5) / f.y_size() as Float,
+                ));
+                f[(x as i32, y as i32)] = func(p);
+            }
+        }
+        let distrib = PiecewiseConstant2D::new_with_bounds(&f, domain);
+        Self { domain, distrib, f }
+    }
+
     pub fn sample(&self, u: Point2f) -> FilterSample {
         let (p, pdf, pi) = self.distrib.sample(u);
-
         if pdf == 0.0 {
             return FilterSample { p, weight: 0.0 };
         }
-
-        let idx = pi.x() as u32 + self.f.x_size();
-        let weight = *self.f.get_linear(idx as usize) / pdf;
+        let idx = pi.x() as usize + pi.y() as usize * self.f.x_size();
+        let weight = self.f.as_slice()[idx] / pdf;
         FilterSample { p, weight }
     }
 }
@@ -38,7 +59,7 @@ pub trait FilterTrait {
     fn radius(&self) -> Vector2f;
     fn evaluate(&self, p: Point2f) -> Float;
     fn integral(&self) -> Float;
-    fn sample(&self, u: Point2f) -> DeviceFilterSample;
+    fn sample(&self, u: Point2f) -> FilterSample;
 }
 
 #[repr(C)]
diff --git a/shared/src/core/image.rs b/shared/src/core/image.rs
index a9025ee..ae1f3c7 100644
--- a/shared/src/core/image.rs
+++ b/shared/src/core/image.rs
@@ -1,9 +1,9 @@
-use crate::Float;
 use crate::core::color::{ColorEncoding, ColorEncodingTrait, LINEAR};
 use crate::core::geometry::{Bounds2f, Point2f, Point2fi, Point2i};
-use crate::utils::Ptr;
 use crate::utils::containers::DeviceArray2D;
 use crate::utils::math::{f16_to_f32_software, lerp, square};
+use crate::Float;
+use crate::GVec;
 use core::hash;
 use core::ops::{Deref, DerefMut};
 use num_traits::Float as NumFloat;
@@ -68,119 +68,305 @@ impl PixelFormat {
 }
 
 #[repr(C)]
-#[derive(Clone, Copy, Debug)]
+#[derive(Clone, Debug)]
 pub struct Pixels {
-    ptr: Ptr<u8>,
+    pixels: GVec<u8>,
     format: PixelFormat,
 }
 
 impl Pixels {
-    pub fn new_u8(ptr: Ptr<u8>) -> Self {
+    pub fn new(data: GVec<u8>, format: PixelFormat) -> Self {
+        Self { data, format }
+    }
+
+    pub fn format(&self) -> PixelFormat {
+        self.format
+    }
+
+    pub fn as_ptr(&self) -> *const u8 {
+        self.data.as_ptr()
+    }
+
+    pub fn len(&self) -> usize {
+        self.data.len()
+    }
+
+    pub fn texel_count(&self) -> usize {
+        self.data.len() / self.format.texel_bytes()
+    }
+
+    pub unsafe fn read_u8(&self, texel_offset: usize) -> u8 {
+        *self.data.as_ptr().add(texel_offset)
+    }
+
+    pub unsafe fn read_f16(&self, texel_offset: usize) -> u16 {
+        let byte_offset = texel_offset * 2;
+        *(self.data.as_ptr().add(byte_offset) as *const u16)
+    }
+
+    pub unsafe fn read_f32(&self, texel_offset: usize) -> f32 {
+        let byte_offset = texel_offset * 4;
+        *(self.data.as_ptr().add(byte_offset) as *const f32)
+    }
+
+    pub unsafe fn read(&self, texel_offset: usize, encoding: &ColorEncoding) -> Float {
+        match self.format {
+            PixelFormat::U8 => encoding.to_linear_scalar(self.read_u8(texel_offset)),
+            PixelFormat::F16 => f16_to_f32_software(self.read_f16(texel_offset)),
+            PixelFormat::F32 => self.read_f32(texel_offset),
+        }
+    }
+
+    pub unsafe fn write_u8(&mut self, texel_offset: usize, val: u8) {
+        *self.data.as_mut_ptr().add(texel_offset) = val;
+    }
+
+    pub unsafe fn write_f16(&mut self, texel_offset: usize, val: u16) {
+        let byte_offset = texel_offset * 2;
+        *(self.data.as_mut_ptr().add(byte_offset) as *mut u16) = val;
+    }
+
+    pub unsafe fn write_f32(&mut self, texel_offset: usize, val: f32) {
+        let byte_offset = texel_offset * 4;
+        *(self.data.as_mut_ptr().add(byte_offset) as *mut f32) = val;
+    }
+
+    pub fn empty(texel_count: usize, format: PixelFormat) -> Self {
+        let byte_count = texel_count * format.texel_bytes();
+        let mut data = gvec_with_capacity(byte_count);
+        data.resize(byte_count, 0u8);
+        Self { data, format }
+    }
+
+    pub fn from_u8_slice(slice: &[u8]) -> Self {
+        let mut data = gvec_with_capacity(slice.len());
+        data.extend_from_slice(slice);
         Self {
-            ptr,
+            data,
             format: PixelFormat::U8,
         }
     }
 
-    pub fn new_f16(ptr: Ptr<u16>) -> Self {
+    pub fn from_f32_slice(slice: &[f32]) -> Self {
+        let byte_len = slice.len() * 4;
+        let mut data = gvec_with_capacity(byte_len);
+        let bytes = unsafe { core::slice::from_raw_parts(slice.as_ptr() as *const u8, byte_len) };
+        data.extend_from_slice(bytes);
         Self {
-            ptr: Ptr::from_raw(ptr.as_raw() as *const u8),
-            format: PixelFormat::F16,
-        }
-    }
-
-    pub fn new_f32(ptr: Ptr<f32>) -> Self {
-        Self {
-            ptr: Ptr::from_raw(ptr.as_raw() as *const u8),
+            data,
             format: PixelFormat::F32,
         }
     }
 
-    unsafe fn read_u8(&self, byte_offset: usize) -> u8 {
-        unsafe { *self.ptr.as_raw().add(byte_offset) }
+    pub fn from_f16_slice(slice: &[u16]) -> Self {
+        let byte_len = slice.len() * 2;
+        let mut data = gvec_with_capacity(byte_len);
+        let bytes = unsafe { core::slice::from_raw_parts(slice.as_ptr() as *const u8, byte_len) };
+        data.extend_from_slice(bytes);
+        Self {
+            data,
+            format: PixelFormat::F16,
+        }
     }
 
-    unsafe fn read_f16(&self, elem_offset: usize) -> u16 {
-        let byte_offset = elem_offset * 2;
-        unsafe { *(self.ptr.as_raw().add(byte_offset) as *const u16) }
+    pub fn as_u8_mut(&mut self) -> &mut [u8] {
+        &mut self.data
     }
 
-    unsafe fn read_f32(&self, elem_offset: usize) -> f32 {
-        let byte_offset = elem_offset * 4;
-        unsafe { *(self.ptr.as_raw().add(byte_offset) as *const f32) }
+    pub fn as_f16_mut(&mut self) -> &mut [u16] {
+        assert_eq!(self.format, PixelFormat::F16);
+        unsafe {
+            core::slice::from_raw_parts_mut(self.data.as_mut_ptr() as *mut u16, self.data.len() / 2)
+        }
     }
 
-    // Unsure if ill need this
-    // pub unsafe fn read(&self, offset: usize) -> Float {
-    //     match self.format {
-    //         PixelFormat::U8 => unsafe { self.read_u8(offset) as Float / 255.0 },
-    //         PixelFormat::F16 => unsafe { f16_to_f32_software(self.read_f16(offset)) },
-    //         PixelFormat::F32 => unsafe { self.read_f32(offset) },
-    //     }
-    // }
+    pub fn as_f32_slice(&self) -> &[f32] {
+        assert_eq!(self.format, PixelFormat::F32);
+        unsafe {
+            core::slice::from_raw_parts(self.data.as_ptr() as *const f32, self.data.len() / 4)
+        }
+    }
+
+    pub fn as_f32_slice_mut(&mut self) -> &mut [f32] {
+        assert_eq!(self.format, PixelFormat::F32);
+        unsafe {
+            core::slice::from_raw_parts_mut(self.data.as_mut_ptr() as *mut f32, self.data.len() / 4)
+        }
+    }
 }
 
-#[repr(C)]
-#[derive(Clone, Copy, Debug)]
-pub struct ImageBase {
+#[derive(Clone, Debug)]
+pub struct Image {
     pub format: PixelFormat,
     pub encoding: ColorEncoding,
     pub resolution: Point2i,
     pub n_channels: i32,
+    pub pixels: Pixels,
 }
 
-impl ImageBase {
+impl Image {
+    pub fn new(
+        format: PixelFormat,
+        resolution: Point2i,
+        n_channels: i32,
+        encoding: ColorEncoding,
+    ) -> Self {
+        let texel_count = (resolution.x() * resolution.y()) as usize * n_channels as usize;
+        Self {
+            format,
+            encoding,
+            resolution,
+            n_channels,
+            pixels: Pixels::empty(texel_count, format),
+        }
+    }
+
+    pub fn from_u8(
+        data: &[u8],
+        resolution: Point2i,
+        n_channels: i32,
+        encoding: ColorEncoding,
+    ) -> Self {
+        let expected = (resolution.x() * resolution.y()) as usize * n_channels as usize;
+        assert_eq!(data.len(), expected, "Pixel data size mismatch");
+        Self {
+            format: PixelFormat::U8,
+            encoding,
+            resolution,
+            n_channels,
+            pixels: Pixels::from_u8_slice(data),
+        }
+    }
+
+    pub fn from_f32(data: &[f32], resolution: Point2i, n_channels: i32) -> Self {
+        let expected = (resolution.x() * resolution.y()) as usize * n_channels as usize;
+        assert_eq!(data.len(), expected, "Pixel data size mismatch");
+        Self {
+            format: PixelFormat::F32,
+            encoding: LINEAR,
+            resolution,
+            n_channels,
+            pixels: Pixels::from_f32_slice(data),
+        }
+    }
+
+    pub fn from_f16(data: &[u16], resolution: Point2i, n_channels: i32) -> Self {
+        let expected = (resolution.x() * resolution.y()) as usize * n_channels as usize;
+        assert_eq!(data.len(), expected, "Pixel data size mismatch");
+        Self {
+            format: PixelFormat::F16,
+            encoding: LINEAR,
+            resolution,
+            n_channels,
+            pixels: Pixels::from_f16_slice(data),
+        }
+    }
+
+    pub fn resolution(&self) -> Point2i {
+        self.resolution
+    }
+
+    pub fn n_channels(&self) -> i32 {
+        self.n_channels
+    }
+
+    pub fn format(&self) -> PixelFormat {
+        self.format
+    }
+
+    pub fn is_valid(&self) -> bool {
+        self.resolution.x() > 0 && self.resolution.y() > 0
+    }
+
+    pub fn pixel_offset(&self, p: Point2i) -> usize {
+        let width = self.resolution.x() as usize;
+        (p.y() as usize * width + p.x() as usize) * self.n_channels as usize
+    }
+
     pub fn remap_pixel_coords(&self, p: &mut Point2i, wrap_mode: WrapMode2D) -> bool {
-        let resolution = self.resolution;
         for i in 0..2 {
-            if p[i] >= 0 && p[i] < resolution[i] {
+            if p[i] >= 0 && p[i] < self.resolution[i] {
                 continue;
             }
             match wrap_mode.uv[i] {
                 WrapMode::Black => return false,
-                WrapMode::Clamp => p[i] = p[i].clamp(0, resolution[i] - 1),
-                WrapMode::Repeat => p[i] = p[i].rem_euclid(resolution[i]),
+                WrapMode::Clamp => p[i] = p[i].clamp(0, self.resolution[i] - 1),
+                WrapMode::Repeat => p[i] = p[i].rem_euclid(self.resolution[i]),
                 WrapMode::OctahedralSphere => {
-                    p[i] = p[i].clamp(0, resolution[i] - 1);
+                    p[i] = p[i].clamp(0, self.resolution[i] - 1);
                 }
             }
         }
         true
     }
-}
 
-#[repr(C)]
-#[derive(Debug, Clone, Copy)]
-pub struct DeviceImage {
-    pub base: ImageBase,
-    pub pixels: Pixels,
-}
-
-impl DeviceImage {
-    pub fn base(&self) -> ImageBase {
-        self.base
+    pub fn get_channel(&self, p: Point2i, c: i32) -> Float {
+        self.get_channel_with_wrap(p, c, WrapMode::Clamp.into())
     }
 
-    pub fn resolution(&self) -> Point2i {
-        self.base.resolution
+    pub fn get_channel_with_wrap(&self, mut p: Point2i, c: i32, wrap_mode: WrapMode2D) -> Float {
+        if !self.remap_pixel_coords(&mut p, wrap_mode) {
+            return 0.0;
+        }
+        let offset = self.pixel_offset(p) + c as usize;
+        unsafe { self.pixels.read(offset, &self.encoding) }
     }
 
-    pub fn is_valid(&self) -> bool {
-        self.resolution().x() > 0 && self.resolution().y() > 0
+    pub fn get_channels_array<const N: usize>(&self, p: Point2i) -> [Float; N] {
+        self.get_channels_array_with_wrap(p, WrapMode::Clamp.into())
     }
 
-    pub fn format(&self) -> PixelFormat {
-        self.base().format
+    pub fn get_channels_array_with_wrap<const N: usize>(
+        &self,
+        mut p: Point2i,
+        wrap_mode: WrapMode2D,
+    ) -> [Float; N] {
+        debug_assert!(N <= self.n_channels as usize);
+        let mut result = [0.0; N];
+        if !self.remap_pixel_coords(&mut p, wrap_mode) {
+            return result;
+        }
+        let offset = self.pixel_offset(p);
+        for i in 0..N {
+            result[i] = unsafe { self.pixels.read(offset + i, &self.encoding) };
+        }
+        result
     }
 
-    pub fn n_channels(&self) -> i32 {
-        self.base().n_channels
+    pub fn get_channels_average(&self, p: Point2i) -> Float {
+        let offset = self.pixel_offset(p);
+        let nc = self.n_channels as usize;
+        let mut sum = 0.0;
+        for i in 0..nc {
+            sum += unsafe { self.pixels.read(offset + i, &self.encoding) };
+        }
+        sum / nc as Float
     }
 
-    pub fn pixel_offset(&self, p: Point2i) -> u32 {
-        let width = self.resolution().x() as u32;
-        let idx = p.y() as u32 * width + p.x() as u32;
-        idx * (self.n_channels() as u32)
+    pub fn set_channel(&mut self, p: Point2i, c: i32, mut value: Float) {
+        if value.is_nan() {
+            value = 0.0;
+        }
+        let res = self.resolution;
+        if p.x() < 0 || p.x() >= res.x() || p.y() < 0 || p.y() >= res.y() {
+            return;
+        }
+        let offset = self.pixel_offset(p) + c as usize;
+        unsafe {
+            match self.format {
+                PixelFormat::U8 => {
+                    self.pixels
+                        .write_u8(offset, self.encoding.from_linear_scalar(value));
+                }
+                PixelFormat::F16 => {
+                    self.pixels
+                        .write_f16(offset, half::f16::from_f32(value).to_bits());
+                }
+                PixelFormat::F32 => {
+                    self.pixels.write_f32(offset, value);
+                }
+            }
+        }
     }
 
     pub fn bilerp_channel(&self, p: Point2f, c: i32) -> Float {
@@ -188,8 +374,8 @@ impl DeviceImage {
     }
 
     pub fn bilerp_channel_with_wrap(&self, p: Point2f, c: i32, wrap_mode: WrapMode2D) -> Float {
-        let x = p.x() * self.resolution().x() as Float - 0.5;
-        let y = p.y() * self.resolution().y() as Float - 0.5;
+        let x = p.x() * self.resolution.x() as Float - 0.5;
+        let y = p.y() * self.resolution.y() as Float - 0.5;
         let xi = x.floor() as i32;
         let yi = y.floor() as i32;
         let dx = x - xi as Float;
@@ -200,51 +386,30 @@ impl DeviceImage {
         let v11 = self.get_channel_with_wrap(Point2i::new(xi + 1, yi + 1), c, wrap_mode);
         lerp(dy, lerp(dx, v00, v10), lerp(dx, v01, v11))
     }
-}
 
-pub trait ImageAccess {
-    fn get_channel_with_wrap(&self, p: Point2i, c: i32, wrap_mode: WrapMode2D) -> Float;
-    fn get_channel(&self, p: Point2i, c: i32) -> Float;
-    fn lookup_nearest_channel_with_wrap(&self, p: Point2f, c: i32, wrap_mode: WrapMode2D) -> Float;
-    fn lookup_nearest_channel(&self, p: Point2f, c: i32) -> Float;
-}
-
-impl ImageAccess for DeviceImage {
-    fn get_channel_with_wrap(&self, mut p: Point2i, c: i32, wrap_mode: WrapMode2D) -> Float {
-        if !self.base.remap_pixel_coords(&mut p, wrap_mode) {
-            return 0.;
-        }
-
-        let offset = (self.pixel_offset(p) + c as u32) as usize;
-        unsafe {
-            match self.pixels.format {
-                PixelFormat::U8 => {
-                    let raw_val = self.pixels.read_u8(offset);
-                    self.base.encoding.to_linear_scalar(raw_val)
+    pub fn has_any_infinite_pixels(&self) -> bool {
+        for y in 0..self.resolution.y() {
+            for x in 0..self.resolution.x() {
+                for c in 0..self.n_channels {
+                    if self.get_channel(Point2i::new(x, y), c).is_infinite() {
+                        return true;
+                    }
                 }
-                PixelFormat::F16 => {
-                    let raw_val = self.pixels.read_f16(offset);
-                    f16_to_f32_software(raw_val)
-                }
-                PixelFormat::F32 => self.pixels.read_f32(offset),
             }
         }
+        false
     }
 
-    fn get_channel(&self, p: Point2i, c: i32) -> Float {
-        self.get_channel_with_wrap(p, c, WrapMode::Clamp.into())
-    }
-
-    fn lookup_nearest_channel_with_wrap(&self, p: Point2f, c: i32, wrap_mode: WrapMode2D) -> Float {
-        let pi = Point2i::new(
-            p.x() as i32 * self.resolution().x(),
-            p.y() as i32 * self.resolution().y(),
-        );
-
-        self.get_channel_with_wrap(pi, c, wrap_mode)
-    }
-
-    fn lookup_nearest_channel(&self, p: Point2f, c: i32) -> Float {
-        self.lookup_nearest_channel_with_wrap(p, c, WrapMode::Clamp.into())
+    pub fn has_any_nan_pixels(&self) -> bool {
+        for y in 0..self.resolution.y() {
+            for x in 0..self.resolution.x() {
+                for c in 0..self.n_channels {
+                    if self.get_channel(Point2i::new(x, y), c).is_nan() {
+                        return true;
+                    }
+                }
+            }
+        }
+        false
     }
 }
diff --git a/shared/src/core/interaction.rs b/shared/src/core/interaction.rs
index 9b70526..13fef6c 100644
--- a/shared/src/core/interaction.rs
+++ b/shared/src/core/interaction.rs
@@ -7,7 +7,7 @@ use crate::core::camera::{Camera, CameraTrait};
 use crate::core::geometry::{
     Normal3f, Point2f, Point3f, Point3fi, Ray, RayDifferential, Vector3f, VectorLike,
 };
-use crate::core::image::DeviceImage;
+use crate::core::image::Image;
 use crate::core::light::{Light, LightTrait};
 use crate::core::material::{
     Material, MaterialEvalContext, MaterialTrait, NormalBumpEvalContext, bump_map, normal_map,
@@ -348,7 +348,7 @@ impl SurfaceInteraction {
         &mut self,
         tex_eval: &UniversalTextureEvaluator,
         displacement: Ptr<GPUFloatTexture>,
-        normal_image: Ptr<DeviceImage>,
+        normal_image: Ptr<Image>,
     ) {
         let ctx = NormalBumpEvalContext::from(&*self);
         let (dpdu, dpdv) = if !displacement.is_null() {
diff --git a/shared/src/core/light.rs b/shared/src/core/light.rs
index a2a1762..c71b6c0 100644
--- a/shared/src/core/light.rs
+++ b/shared/src/core/light.rs
@@ -3,7 +3,6 @@ use crate::core::geometry::{
     Bounds2f, Bounds3f, DirectionCone, Normal3f, Point2f, Point2i, Point3f, Point3fi, Ray,
     Vector3f, VectorLike, cos_theta,
 };
-use crate::core::image::DeviceImage;
 use crate::core::interaction::{
     Interaction, InteractionBase, InteractionTrait, MediumInteraction, SimpleInteraction,
     SurfaceInteraction,
@@ -17,7 +16,6 @@ use crate::spectra::{
 };
 use crate::utils::Transform;
 use crate::utils::math::{equal_area_sphere_to_square, radians, safe_sqrt, smooth_step, square};
-use crate::utils::sampling::DevicePiecewiseConstant2D;
 use crate::{Float, PI};
 use bitflags::bitflags;
 
diff --git a/shared/src/core/material.rs b/shared/src/core/material.rs
index 011d6bb..99cec02 100644
--- a/shared/src/core/material.rs
+++ b/shared/src/core/material.rs
@@ -10,7 +10,7 @@ use crate::core::bsdf::BSDF;
 use crate::core::bssrdf::BSSRDF;
 use crate::core::bxdf::BxDF;
 use crate::core::geometry::{Frame, Normal3f, Point2f, Point3f, Vector2f, Vector3f, VectorLike};
-use crate::core::image::{DeviceImage, WrapMode, WrapMode2D};
+use crate::core::image::{Image, WrapMode, WrapMode2D};
 use crate::core::interaction::{Interaction, InteractionTrait, ShadingGeom, SurfaceInteraction};
 use crate::core::scattering::TrowbridgeReitzDistribution;
 use crate::core::spectrum::{Spectrum, SpectrumTrait};
@@ -103,7 +103,7 @@ impl From<&NormalBumpEvalContext> for TextureEvalContext {
     }
 }
 
-pub fn normal_map(normal_map: &DeviceImage, ctx: &NormalBumpEvalContext) -> (Vector3f, Vector3f) {
+pub fn normal_map(normal_map: &Image, ctx: &NormalBumpEvalContext) -> (Vector3f, Vector3f) {
     let wrap = WrapMode2D::from(WrapMode::Repeat);
     let uv = Point2f::new(ctx.uv[0], 1. - ctx.uv[1]);
     let r = normal_map.bilerp_channel_with_wrap(uv, 0, wrap);
@@ -173,7 +173,7 @@ pub trait MaterialTrait {
     ) -> Option<BSSRDF>;
 
     fn can_evaluate_textures(&self, tex_eval: &dyn TextureEvaluator) -> bool;
-    fn get_normal_map(&self) -> Option<&DeviceImage>;
+    fn get_normal_map(&self) -> Option<&Image>;
     fn get_displacement(&self) -> Ptr<GPUFloatTexture>;
     fn has_subsurface_scattering(&self) -> bool;
 }
diff --git a/shared/src/core/pbrt.rs b/shared/src/core/pbrt.rs
index b276c5a..d63f343 100644
--- a/shared/src/core/pbrt.rs
+++ b/shared/src/core/pbrt.rs
@@ -2,7 +2,6 @@ use crate::core::geometry::Lerp;
 use core::ops::{Add, Mul};
 use num_traits::{Num, PrimInt};
 
-use crate::core::image::DeviceImage;
 use crate::core::light::LightTrait;
 use crate::core::shape::Shape;
 use crate::core::texture::GPUFloatTexture;
diff --git a/shared/src/core/primitive.rs b/shared/src/core/primitive.rs
index 72b3476..98df32f 100644
--- a/shared/src/core/primitive.rs
+++ b/shared/src/core/primitive.rs
@@ -1,5 +1,5 @@
 use crate::core::geometry::{Bounds3f, Ray};
-use crate::core::aggregates::DeviceBVHAggregate;
+use crate::core::aggregates::BVHAggregate;
 use crate::core::interaction::{Interaction, InteractionTrait, SurfaceInteraction};
 use crate::core::light::Light;
 use crate::core::material::Material;
@@ -213,6 +213,6 @@ pub enum Primitive {
     Geometric(GeometricPrimitive),
     Transformed(TransformedPrimitive),
     Animated(AnimatedPrimitive),
-    BVH(DeviceBVHAggregate),
+    BVH(BVHAggregate),
     KdTree(KdTreeAggregate),
 }
diff --git a/shared/src/core/sampler.rs b/shared/src/core/sampler.rs
index 14e077c..760fb6d 100644
--- a/shared/src/core/sampler.rs
+++ b/shared/src/core/sampler.rs
@@ -1,17 +1,17 @@
 use crate::core::filter::FilterTrait;
 use crate::core::geometry::{Bounds2f, Point2f, Point2i, Vector2f};
 use crate::core::pbrt::{Float, ONE_MINUS_EPSILON, PI, PI_OVER_2, PI_OVER_4};
-use crate::utils::Ptr;
-use crate::utils::containers::DeviceArray2D;
 use crate::utils::math::{
-    BinaryPermuteScrambler, DeviceDigitPermutation, FastOwenScrambler, NoRandomizer, OwenScrambler,
-    PRIME_TABLE_SIZE, Scrambler, clamp, encode_morton_2, inverse_radical_inverse, lerp, log2_int,
+    clamp, encode_morton_2, inverse_radical_inverse, lerp, log2_int,
     owen_scrambled_radical_inverse, permutation_element, radical_inverse, round_up_pow2,
-    scrambled_radical_inverse, sobol_interval_to_index, sobol_sample,
+    scrambled_radical_inverse, sobol_interval_to_index, sobol_sample, BinaryPermuteScrambler,
+    DeviceDigitPermutation, FastOwenScrambler, NoRandomizer, OwenScrambler, Scrambler,
+    PRIME_TABLE_SIZE,
 };
 use crate::utils::rng::Rng;
 use crate::utils::sobol::N_SOBOL_DIMENSIONS;
 use crate::utils::{hash::*, sobol};
+use crate::{gvec, GVec, Ptr};
 use enum_dispatch::enum_dispatch;
 
 #[repr(C)]
@@ -100,7 +100,7 @@ pub struct HaltonSampler {
     pub mult_inverse: [u64; 2],
     pub halton_index: u64,
     pub dim: u32,
-    pub digit_permutations: Ptr<DeviceDigitPermutation>,
+    pub digit_permutations: GVec<DeviceDigitPermutation>,
 }
 
 #[allow(clippy::derivable_impls)]
@@ -114,7 +114,7 @@ impl Default for HaltonSampler {
             mult_inverse: [0; 2],
             halton_index: 0,
             dim: 0,
-            digit_permutations: Ptr::default(),
+            digit_permutations: gvec(),
         }
     }
 }
diff --git a/shared/src/filters/boxf.rs b/shared/src/filters/boxf.rs
index 413cb8d..bee4080 100644
--- a/shared/src/filters/boxf.rs
+++ b/shared/src/filters/boxf.rs
@@ -1,5 +1,5 @@
 use crate::Float;
-use crate::core::filter::{DeviceFilterSample, FilterTrait};
+use crate::core::filter::{FilterSample, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use crate::utils::math::lerp;
 
@@ -31,7 +31,7 @@ impl FilterTrait for BoxFilter {
         (2.0 * self.radius.x()) * (2.0 * self.radius.y())
     }
 
-    fn sample(&self, u: Point2f) -> DeviceFilterSample {
+    fn sample(&self, u: Point2f) -> FilterSample {
         let p = Point2f::new(
             lerp(u[0], -self.radius.x(), self.radius.x()),
             lerp(u[1], -self.radius.y(), self.radius.y()),
diff --git a/shared/src/filters/gaussian.rs b/shared/src/filters/gaussian.rs
index da81189..e10a494 100644
--- a/shared/src/filters/gaussian.rs
+++ b/shared/src/filters/gaussian.rs
@@ -1,7 +1,7 @@
-use crate::Float;
-use crate::core::filter::{DeviceFilterSample, FilterSampler, FilterTrait};
+use crate::core::filter::{FilterSample, FilterSampler, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use crate::utils::math::{gaussian, gaussian_integral};
+use crate::Float;
 
 #[repr(C)]
 #[derive(Clone, Debug, Copy)]
@@ -13,6 +13,25 @@ pub struct GaussianFilter {
     pub sampler: FilterSampler,
 }
 
+impl GaussianFilter {
+    pub fn new(radius: Vector2f, sigma: Float) -> Self {
+        let exp_x = gaussian(radius.x(), 0.0, sigma);
+        let exp_y = gaussian(radius.y(), 0.0, sigma);
+        let sampler = FilterSampler::new(radius, move |p: Point2f| {
+            let gx = (gaussian(p.x(), 0.0, sigma) - exp_x).max(0.0);
+            let gy = (gaussian(p.y(), 0.0, sigma) - exp_y).max(0.0);
+            gx * gy
+        });
+        Self {
+            radius,
+            sigma,
+            exp_x,
+            exp_y,
+            sampler,
+        }
+    }
+}
+
 impl FilterTrait for GaussianFilter {
     fn radius(&self) -> Vector2f {
         self.radius
@@ -30,7 +49,7 @@ impl FilterTrait for GaussianFilter {
                 - 2.0 * self.radius.y() * self.exp_y)
     }
 
-    fn sample(&self, u: Point2f) -> DeviceFilterSample {
+    fn sample(&self, u: Point2f) -> FilterSample {
         self.sampler.sample(u)
     }
 }
diff --git a/shared/src/filters/lanczos.rs b/shared/src/filters/lanczos.rs
index a22db3d..d763f17 100644
--- a/shared/src/filters/lanczos.rs
+++ b/shared/src/filters/lanczos.rs
@@ -1,17 +1,30 @@
-use crate::Float;
-use crate::core::filter::{DeviceFilterSample, FilterSampler, FilterTrait};
+use crate::core::filter::{DeviceFilterSampler, FilterSample, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use crate::utils::math::{lerp, windowed_sinc};
+use crate::Float;
 
 #[repr(C)]
 #[derive(Clone, Debug, Copy)]
 pub struct LanczosSincFilter {
     pub radius: Vector2f,
     pub tau: Float,
-    pub sampler: FilterSampler,
+    pub sampler: DeviceFilterSampler,
     pub integral: Float,
 }
 
+impl LanczosSincFilter {
+    pub fn new(radius: Vector2f, tau: Float) -> Self {
+        let sampler = FilterSampler::new(radius, move |p: Point2f| {
+            windowed_sinc(p.x(), radius.x(), tau) * windowed_sinc(p.y(), radius.y(), tau)
+        });
+        Self {
+            radius,
+            tau,
+            sampler,
+        }
+    }
+}
+
 impl FilterTrait for LanczosSincFilter {
     fn radius(&self) -> Vector2f {
         self.radius
@@ -26,7 +39,7 @@ impl FilterTrait for LanczosSincFilter {
         self.integral
     }
 
-    fn sample(&self, u: Point2f) -> DeviceFilterSample {
+    fn sample(&self, u: Point2f) -> FilterSample {
         self.sampler.sample(u)
     }
 }
diff --git a/shared/src/filters/mitchell.rs b/shared/src/filters/mitchell.rs
index c17ceb0..4f19b4a 100644
--- a/shared/src/filters/mitchell.rs
+++ b/shared/src/filters/mitchell.rs
@@ -1,6 +1,6 @@
-use crate::Float;
-use crate::core::filter::{DeviceFilterSample, FilterSampler, FilterTrait};
+use crate::core::filter::{FilterSampler, FilterSample, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
+use crate::Float;
 use num_traits::Float as NumFloat;
 
 #[repr(C)]
@@ -9,7 +9,7 @@ pub struct MitchellFilter {
     pub radius: Vector2f,
     pub b: Float,
     pub c: Float,
-    pub sampler: DeviceFilterSampler,
+    pub sampler: FilterSampler,
 }
 
 impl MitchellFilter {
@@ -34,6 +34,18 @@ impl MitchellFilter {
     fn mitchell_1d(&self, x: Float) -> Float {
         Self::mitchell_1d_eval(self.b, self.c, x)
     }
+
+    pub fn new(radius: Vector2f, b: Float, c: Float) -> Self {
+        let sampler = FilterSampler::new(radius, move |p: Point2f| {
+            mitchell_1d(p.x() / radius.x(), b, c) * mitchell_1d(p.y() / radius.y(), b, c)
+        });
+        Self {
+            radius,
+            b,
+            c,
+            sampler,
+        }
+    }
 }
 
 impl FilterTrait for MitchellFilter {
@@ -50,7 +62,7 @@ impl FilterTrait for MitchellFilter {
         self.radius.x() * self.radius.y() / 4.0
     }
 
-    fn sample(&self, u: Point2f) -> DeviceFilterSample {
+    fn sample(&self, u: Point2f) -> FilterSample {
         self.sampler.sample(u)
     }
 }
diff --git a/shared/src/filters/triangle.rs b/shared/src/filters/triangle.rs
index b014fce..1572ca5 100644
--- a/shared/src/filters/triangle.rs
+++ b/shared/src/filters/triangle.rs
@@ -1,5 +1,5 @@
 use crate::Float;
-use crate::core::filter::{DeviceFilterSample, FilterTrait};
+use crate::core::filter::{FilterSample, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use crate::utils::math::sample_tent;
 use num_traits::Float as NumFloat;
@@ -29,11 +29,11 @@ impl FilterTrait for TriangleFilter {
         self.radius.x().powi(2) * self.radius.y().powi(2)
     }
 
-    fn sample(&self, u: Point2f) -> DeviceFilterSample {
+    fn sample(&self, u: Point2f) -> FilterSample {
         let p = Point2f::new(
             sample_tent(u[0], self.radius.x()),
             sample_tent(u[1], self.radius.y()),
         );
-        DeviceFilterSample { p, weight: 1.0 }
+        FilterSample { p, weight: 1.0 }
     }
 }
diff --git a/shared/src/lib.rs b/shared/src/lib.rs
index d36af1a..68d6b8e 100644
--- a/shared/src/lib.rs
+++ b/shared/src/lib.rs
@@ -17,5 +17,5 @@ pub mod textures;
 pub mod utils;
 
 pub use core::pbrt::*;
-pub use utils::PBRTOptions;
-pub use utils::ptr::Ptr;
+pub use utils::alloc::{gbox, gvec, gvec_from_slice, gvec_with_capacity, GVec, Gbox};
+pub use utils::{Transform, Ptr, Array2D, PBRTOptions};
diff --git a/shared/src/lights/diffuse.rs b/shared/src/lights/diffuse.rs
index 6d09aa3..22f4341 100644
--- a/shared/src/lights/diffuse.rs
+++ b/shared/src/lights/diffuse.rs
@@ -1,6 +1,6 @@
 use crate::core::color::{RGB, XYZ};
 use crate::core::geometry::*;
-use crate::core::image::{DeviceImage, ImageAccess};
+use crate::core::image::{Image, ImageAccess};
 use crate::core::interaction::{
     Interaction, InteractionTrait, MediumInteraction, SurfaceInteraction,
 };
@@ -27,7 +27,7 @@ pub struct DiffuseAreaLight {
     pub alpha: Ptr<GPUFloatTexture>,
     pub colorspace: Ptr<RGBColorSpace>,
     pub lemit: Ptr<DenselySampledSpectrum>,
-    pub image: Ptr<DeviceImage>,
+    pub image: Ptr<Image>,
     pub area: Float,
     pub two_sided: bool,
     pub scale: Float,
diff --git a/shared/src/lights/goniometric.rs b/shared/src/lights/goniometric.rs
index 30b446d..995a8f3 100644
--- a/shared/src/lights/goniometric.rs
+++ b/shared/src/lights/goniometric.rs
@@ -1,5 +1,5 @@
 use crate::core::geometry::{Bounds3f, Normal3f, Point2f, Point2i, Point3f, Ray, Vector3f};
-use crate::core::image::{DeviceImage, ImageAccess};
+use crate::core::image::{Image, ImageAccess};
 use crate::core::light::{
     LightBase, LightBounds, LightLiSample, LightSampleContext, LightTrait, LightType,
 };
@@ -7,7 +7,7 @@ use crate::core::medium::MediumInterface;
 use crate::core::spectrum::{Spectrum, SpectrumTrait};
 use crate::spectra::{DenselySampledSpectrum, SampledSpectrum, SampledWavelengths};
 use crate::utils::math::equal_area_sphere_to_square;
-use crate::utils::sampling::DevicePiecewiseConstant2D;
+use crate::utils::sampling::PiecewiseConstant2D;
 use crate::utils::{Ptr, Transform};
 use crate::{Float, PI};
 
@@ -16,8 +16,8 @@ pub struct GoniometricLight {
     pub base: LightBase,
     pub iemit: Ptr<DenselySampledSpectrum>,
     pub scale: Float,
-    pub image: Ptr<DeviceImage>,
-    pub distrib: Ptr<DevicePiecewiseConstant2D>,
+    pub image: Ptr<Image>,
+    pub distrib: Ptr<PiecewiseConstant2D>,
 }
 
 impl GoniometricLight {
diff --git a/shared/src/lights/infinite.rs b/shared/src/lights/infinite.rs
index 779a159..c46f138 100644
--- a/shared/src/lights/infinite.rs
+++ b/shared/src/lights/infinite.rs
@@ -3,7 +3,7 @@ use crate::core::geometry::{
     Bounds2f, Bounds3f, Normal3f, Point2f, Point2i, Point3f, Ray, Vector2f, Vector3f,
 };
 use crate::core::geometry::{Frame, VectorLike};
-use crate::core::image::{DeviceImage, ImageAccess, PixelFormat, WrapMode};
+use crate::core::image::{Image, ImageAccess, PixelFormat, WrapMode};
 use crate::core::interaction::InteractionBase;
 use crate::core::interaction::{Interaction, SimpleInteraction};
 use crate::core::light::{
@@ -15,7 +15,7 @@ use crate::spectra::{DenselySampledSpectrum, SampledSpectrum, SampledWavelengths
 use crate::spectra::{RGBColorSpace, RGBIlluminantSpectrum};
 use crate::utils::math::{clamp, equal_area_sphere_to_square, equal_area_square_to_sphere, square};
 use crate::utils::sampling::{
-    AliasTable, DevicePiecewiseConstant2D, DeviceWindowedPiecewiseConstant2D,
+    AliasTable, DevicePiecewiseConstant2D, WindowedPiecewiseConstant2D,
     sample_uniform_sphere, uniform_sphere_pdf,
 };
 use crate::utils::{Ptr, Transform};
@@ -114,7 +114,7 @@ impl LightTrait for UniformInfiniteLight {
 #[derive(Clone, Copy, Debug)]
 pub struct ImageInfiniteLight {
     pub base: LightBase,
-    pub image: Ptr<DeviceImage>,
+    pub image: Ptr<Image>,
     pub image_color_space: Ptr<RGBColorSpace>,
     pub distrib: Ptr<DevicePiecewiseConstant2D>,
     pub compensated_distrib: Ptr<DevicePiecewiseConstant2D>,
@@ -250,12 +250,12 @@ impl LightTrait for ImageInfiniteLight {
 #[derive(Debug, Copy, Clone)]
 pub struct PortalInfiniteLight {
     pub base: LightBase,
-    pub image: Ptr<DeviceImage>,
+    pub image: Ptr<Image>,
     pub image_color_space: Ptr<RGBColorSpace>,
     pub scale: Float,
     pub portal: [Point3f; 4],
     pub portal_frame: Frame,
-    pub distribution: DeviceWindowedPiecewiseConstant2D,
+    pub distribution: WindowedPiecewiseConstant2D,
     pub scene_center: Point3f,
     pub scene_radius: Float,
 }
diff --git a/shared/src/materials/coated.rs b/shared/src/materials/coated.rs
index 6152acb..8b21457 100644
--- a/shared/src/materials/coated.rs
+++ b/shared/src/materials/coated.rs
@@ -4,7 +4,7 @@ use crate::bxdfs::{
 use crate::core::bsdf::BSDF;
 use crate::core::bssrdf::BSSRDF;
 use crate::core::bxdf::BxDF;
-use crate::core::image::DeviceImage;
+use crate::core::image::Image;
 use crate::core::material::{Material, MaterialEvalContext, MaterialTrait};
 use crate::core::scattering::TrowbridgeReitzDistribution;
 use crate::core::spectrum::{Spectrum, SpectrumTrait};
@@ -16,7 +16,7 @@ use crate::utils::math::clamp;
 #[repr(C)]
 #[derive(Clone, Copy, Debug)]
 pub struct CoatedDiffuseMaterial {
-    pub normal_map: Ptr<DeviceImage>,
+    pub normal_map: Ptr<Image>,
     pub displacement: Ptr<GPUFloatTexture>,
     pub reflectance: Ptr<GPUSpectrumTexture>,
     pub albedo: Ptr<GPUSpectrumTexture>,
@@ -42,7 +42,7 @@ impl CoatedDiffuseMaterial {
         g: Ptr<GPUFloatTexture>,
         eta: Ptr<Spectrum>,
         displacement: Ptr<GPUFloatTexture>,
-        normal_map: Ptr<DeviceImage>,
+        normal_map: Ptr<Image>,
         remap_roughness: bool,
         max_depth: u32,
         n_samples: u32,
@@ -137,7 +137,7 @@ impl MaterialTrait for CoatedDiffuseMaterial {
         )
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         Some(&*self.normal_map)
     }
 
@@ -153,7 +153,7 @@ impl MaterialTrait for CoatedDiffuseMaterial {
 #[repr(C)]
 #[derive(Clone, Copy, Debug)]
 pub struct CoatedConductorMaterial {
-    normal_map: Ptr<DeviceImage>,
+    normal_map: Ptr<Image>,
     displacement: Ptr<GPUFloatTexture>,
     interface_uroughness: Ptr<GPUFloatTexture>,
     interface_vroughness: Ptr<GPUFloatTexture>,
@@ -175,7 +175,7 @@ pub struct CoatedConductorMaterial {
 impl CoatedConductorMaterial {
     #[allow(clippy::too_many_arguments)]
     pub fn new(
-        normal_map: Ptr<DeviceImage>,
+        normal_map: Ptr<Image>,
         displacement: Ptr<GPUFloatTexture>,
         interface_uroughness: Ptr<GPUFloatTexture>,
         interface_vroughness: Ptr<GPUFloatTexture>,
@@ -333,7 +333,7 @@ impl MaterialTrait for CoatedConductorMaterial {
         tex_eval.can_evaluate(&float_textures, &spectrum_textures)
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         Some(&*self.normal_map)
     }
 
diff --git a/shared/src/materials/complex.rs b/shared/src/materials/complex.rs
index f669d6b..3f909ae 100644
--- a/shared/src/materials/complex.rs
+++ b/shared/src/materials/complex.rs
@@ -6,7 +6,7 @@ use crate::bxdfs::{
 use crate::core::bsdf::BSDF;
 use crate::core::bssrdf::{BSSRDF, BSSRDFTable};
 use crate::core::bxdf::BxDF;
-use crate::core::image::DeviceImage;
+use crate::core::image::Image;
 use crate::core::material::{Material, MaterialEvalContext, MaterialTrait};
 use crate::core::scattering::TrowbridgeReitzDistribution;
 use crate::core::spectrum::{Spectrum, SpectrumTrait};
@@ -77,7 +77,7 @@ impl MaterialTrait for HairMaterial {
         todo!()
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         todo!()
     }
 
@@ -94,7 +94,7 @@ impl MaterialTrait for HairMaterial {
 #[derive(Clone, Copy, Debug)]
 pub struct MeasuredMaterial {
     pub displacement: Ptr<GPUFloatTexture>,
-    pub normal_map: Ptr<DeviceImage>,
+    pub normal_map: Ptr<Image>,
     pub brdf: Ptr<MeasuredBxDFData>,
 }
 
@@ -122,7 +122,7 @@ impl MaterialTrait for MeasuredMaterial {
         true
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         Some(&*self.normal_map)
     }
 
@@ -138,7 +138,7 @@ impl MaterialTrait for MeasuredMaterial {
 #[repr(C)]
 #[derive(Clone, Copy, Debug)]
 pub struct SubsurfaceMaterial {
-    pub normal_map: Ptr<DeviceImage>,
+    pub normal_map: Ptr<Image>,
     pub displacement: Ptr<GPUFloatTexture>,
     pub sigma_a: Ptr<GPUSpectrumTexture>,
     pub sigma_s: Ptr<GPUSpectrumMixTexture>,
@@ -174,7 +174,7 @@ impl MaterialTrait for SubsurfaceMaterial {
         todo!()
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         todo!()
     }
 
diff --git a/shared/src/materials/conductor.rs b/shared/src/materials/conductor.rs
index 2353384..26792e3 100644
--- a/shared/src/materials/conductor.rs
+++ b/shared/src/materials/conductor.rs
@@ -4,7 +4,7 @@ use crate::bxdfs::{
 use crate::core::bsdf::BSDF;
 use crate::core::bssrdf::BSSRDF;
 use crate::core::bxdf::BxDF;
-use crate::core::image::DeviceImage;
+use crate::core::image::Image;
 use crate::core::material::{Material, MaterialEvalContext, MaterialTrait};
 use crate::core::scattering::TrowbridgeReitzDistribution;
 use crate::core::spectrum::{Spectrum, SpectrumTrait};
@@ -23,7 +23,7 @@ pub struct ConductorMaterial {
     pub u_roughness: Ptr<GPUFloatTexture>,
     pub v_roughness: Ptr<GPUFloatTexture>,
     pub remap_roughness: bool,
-    pub normal_map: Ptr<DeviceImage>,
+    pub normal_map: Ptr<Image>,
 }
 
 impl MaterialTrait for ConductorMaterial {
@@ -50,7 +50,7 @@ impl MaterialTrait for ConductorMaterial {
         )
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         todo!()
     }
 
diff --git a/shared/src/materials/dielectric.rs b/shared/src/materials/dielectric.rs
index 71be09d..5826c75 100644
--- a/shared/src/materials/dielectric.rs
+++ b/shared/src/materials/dielectric.rs
@@ -4,7 +4,7 @@ use crate::bxdfs::{
 use crate::core::bsdf::BSDF;
 use crate::core::bssrdf::BSSRDF;
 use crate::core::bxdf::BxDF;
-use crate::core::image::DeviceImage;
+use crate::core::image::Image;
 use crate::core::material::{Material, MaterialEvalContext, MaterialTrait};
 use crate::core::scattering::TrowbridgeReitzDistribution;
 use crate::core::spectrum::{Spectrum, SpectrumTrait};
@@ -16,7 +16,7 @@ use crate::utils::math::clamp;
 #[repr(C)]
 #[derive(Clone, Copy, Debug)]
 pub struct DielectricMaterial {
-    pub normal_map: Ptr<DeviceImage>,
+    pub normal_map: Ptr<Image>,
     pub displacement: Ptr<GPUFloatTexture>,
     pub u_roughness: Ptr<GPUFloatTexture>,
     pub v_roughness: Ptr<GPUFloatTexture>,
@@ -67,7 +67,7 @@ impl MaterialTrait for DielectricMaterial {
         tex_eval.can_evaluate(&[self.u_roughness, self.v_roughness], &[])
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         Some(&*self.normal_map)
     }
 
@@ -84,7 +84,7 @@ impl MaterialTrait for DielectricMaterial {
 #[derive(Clone, Copy, Debug)]
 pub struct ThinDielectricMaterial {
     pub displacement: Ptr<GPUFloatTexture>,
-    pub normal_map: Ptr<DeviceImage>,
+    pub normal_map: Ptr<Image>,
     pub eta: Ptr<Spectrum>,
 }
 
@@ -110,7 +110,7 @@ impl MaterialTrait for ThinDielectricMaterial {
         true
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         Some(&*self.normal_map)
     }
 
diff --git a/shared/src/materials/mix.rs b/shared/src/materials/mix.rs
index 424b62d..5e1da28 100644
--- a/shared/src/materials/mix.rs
+++ b/shared/src/materials/mix.rs
@@ -4,7 +4,7 @@ use crate::bxdfs::{
 use crate::core::bsdf::BSDF;
 use crate::core::bssrdf::BSSRDF;
 use crate::core::bxdf::BxDF;
-use crate::core::image::DeviceImage;
+use crate::core::image::Image;
 use crate::core::material::{Material, MaterialEvalContext, MaterialTrait};
 use crate::core::scattering::TrowbridgeReitzDistribution;
 use crate::core::spectrum::{Spectrum, SpectrumTrait};
@@ -69,7 +69,7 @@ impl MaterialTrait for MixMaterial {
         tex_eval.can_evaluate(&[self.amount], &[])
     }
 
-    fn get_normal_map(&self) -> Option<&DeviceImage> {
+    fn get_normal_map(&self) -> Option<&Image> {
         None
     }
 
diff --git a/shared/src/shapes/bilinear.rs b/shared/src/shapes/bilinear.rs
index 53eb6b5..a9269de 100644
--- a/shared/src/shapes/bilinear.rs
+++ b/shared/src/shapes/bilinear.rs
@@ -1,17 +1,16 @@
 use crate::core::geometry::{
-    Bounds3f, DirectionCone, Normal, Normal3f, Point2f, Point3f, Point3fi, Ray, Tuple, Vector3f,
-    VectorLike, spherical_quad_area,
+    spherical_quad_area, Bounds3f, DirectionCone, Normal, Normal3f, Point2f, Point3f, Point3fi,
+    Ray, Tuple, Vector3f, VectorLike,
 };
 use crate::core::interaction::{Interaction, InteractionTrait, SurfaceInteraction};
-use crate::core::pbrt::{Float, gamma};
+use crate::core::pbrt::{gamma, Float};
 use crate::core::shape::{Shape, ShapeIntersection, ShapeSample, ShapeSampleContext, ShapeTrait};
-use crate::utils::Ptr;
-use crate::utils::Transform;
-use crate::utils::math::{SquareMatrix, clamp, difference_of_products, lerp, quadratic};
-use crate::utils::mesh::DeviceBilinearPatchMesh;
+use crate::shapes::mesh::BilinearPatchMesh;
+use crate::utils::math::{clamp, difference_of_products, lerp, quadratic, SquareMatrix};
 use crate::utils::sampling::{
     bilinear_pdf, invert_spherical_rectangle_sample, sample_bilinear, sample_spherical_rectangle,
 };
+use crate::{GVec, Ptr, Transform};
 use core::ops::Add;
 
 #[repr(C)]
@@ -47,7 +46,7 @@ impl BilinearIntersection {
 #[repr(C)]
 #[derive(Debug, Clone, Copy)]
 pub struct BilinearPatchShape {
-    pub mesh: Ptr<DeviceBilinearPatchMesh>,
+    pub mesh: Ptr<BilinearPatchMesh>,
     pub blp_index: i32,
     pub area: Float,
     pub rectangle: bool,
@@ -55,7 +54,7 @@ pub struct BilinearPatchShape {
 
 impl BilinearPatchShape {
     pub const MIN_SPHERICAL_SAMPLE_AREA: Float = 1e-4;
-    fn mesh(&self) -> Ptr<DeviceBilinearPatchMesh> {
+    fn mesh(&self) -> Ptr<BilinearPatchMesh> {
         self.mesh
     }
 
@@ -118,7 +117,7 @@ impl BilinearPatchShape {
     }
 
     #[cfg(not(target_os = "cuda"))]
-    pub fn new(mesh: Ptr<DeviceBilinearPatchMesh>, blp_index: i32) -> Self {
+    pub fn new(mesh: Ptr<BilinearPatchMesh>, blp_index: i32) -> Self {
         let mut bp = BilinearPatchShape {
             mesh,
             blp_index,
@@ -456,7 +455,11 @@ impl BilinearPatchShape {
 
         ss.pdf *= dist_sq / abs_dot;
 
-        if ss.pdf.is_infinite() { None } else { Some(ss) }
+        if ss.pdf.is_infinite() {
+            None
+        } else {
+            Some(ss)
+        }
     }
 
     fn sample_parametric_coords(&self, corners: &[Point3f; 4], u: Point2f) -> (Point2f, Float) {
@@ -723,7 +726,11 @@ impl ShapeTrait for BilinearPatchShape {
 
         let (_, dpdu, dpdv) = self.calculate_base_derivatives(&corners, uv);
         let cross = dpdu.cross(dpdv).norm();
-        if cross == 0. { 0. } else { param_pdf / cross }
+        if cross == 0. {
+            0.
+        } else {
+            param_pdf / cross
+        }
     }
 
     #[inline]
@@ -754,7 +761,11 @@ impl ShapeTrait for BilinearPatchShape {
                 return 0.;
             }
             let pdf = isect_pdf * distsq / absdot;
-            if pdf.is_infinite() { 0. } else { pdf }
+            if pdf.is_infinite() {
+                0.
+            } else {
+                pdf
+            }
         } else {
             let mut pdf = 1. / spherical_quad_area(v00, v10, v01, v11);
             if ctx.ns != Normal3f::zero() {
diff --git a/shared/src/shapes/mesh.rs b/shared/src/shapes/mesh.rs
index cb8aae4..280b703 100644
--- a/shared/src/shapes/mesh.rs
+++ b/shared/src/shapes/mesh.rs
@@ -1,5 +1,5 @@
 use crate::core::geometry::{Normal3f, Point2f, Point3f, Vector3f};
-use crate::utils::sampling::DevicePiecewiseConstant2D;
+use crate::utils::sampling::PiecewiseConstant2D;
 use crate::utils::Transform;
 use crate::{Float, Gvec};
 
diff --git a/shared/src/shapes/mod.rs b/shared/src/shapes/mod.rs
index 16e369f..00be086 100644
--- a/shared/src/shapes/mod.rs
+++ b/shared/src/shapes/mod.rs
@@ -4,6 +4,7 @@ pub mod cylinder;
 pub mod disk;
 pub mod sphere;
 pub mod triangle;
+pub mod mesh;
 
 pub use bilinear::*;
 pub use curves::*;
@@ -11,3 +12,4 @@ pub use cylinder::*;
 pub use disk::*;
 pub use sphere::*;
 pub use triangle::*;
+pub use mesh::*;
diff --git a/shared/src/shapes/triangle.rs b/shared/src/shapes/triangle.rs
index 8a4d543..aa1d33f 100644
--- a/shared/src/shapes/triangle.rs
+++ b/shared/src/shapes/triangle.rs
@@ -1,21 +1,19 @@
-use crate::Float;
+use crate::core::geometry::{spherical_triangle_area, SqrtExt, Tuple, VectorLike};
 use crate::core::geometry::{
     Bounds3f, DirectionCone, Normal, Normal3f, Point2f, Point3f, Point3fi, Ray, Vector2f, Vector3,
     Vector3f,
 };
-use crate::core::geometry::{SqrtExt, Tuple, VectorLike, spherical_triangle_area};
 use crate::core::interaction::{
     Interaction, InteractionBase, InteractionTrait, SimpleInteraction, SurfaceInteraction,
 };
-use crate::core::pbrt::gamma;
 use crate::core::shape::{ShapeIntersection, ShapeSample, ShapeSampleContext, ShapeTrait};
-use crate::utils::Ptr;
+use crate::shapes::mesh::TriangleMesh;
 use crate::utils::math::{difference_of_products, square};
-use crate::utils::mesh::DeviceTriangleMesh;
 use crate::utils::sampling::{
     bilinear_pdf, invert_spherical_triangle_sample, sample_bilinear, sample_spherical_triangle,
     sample_uniform_triangle,
 };
+use crate::{gamma, Float, GVec, Ptr};
 
 #[repr(C)]
 #[derive(Debug, Clone, Copy)]
@@ -35,7 +33,7 @@ impl TriangleIntersection {
 #[repr(C)]
 #[derive(Clone, Copy, Debug)]
 pub struct TriangleShape {
-    pub mesh: Ptr<DeviceTriangleMesh>,
+    pub mesh: Ptr<TriangleMesh>,
     pub tri_index: i32,
 }
 
@@ -43,80 +41,45 @@ impl TriangleShape {
     pub const MIN_SPHERICAL_SAMPLE_AREA: Float = 3e-4;
     pub const MAX_SPHERICAL_SAMPLE_AREA: Float = 6.22;
 
-    #[inline(always)]
     fn get_vertex_indices(&self) -> [usize; 3] {
-        unsafe {
-            let base_ptr = self.mesh.vertex_indices.add((self.tri_index as usize) * 3);
-            [
-                *base_ptr.add(0) as usize,
-                *base_ptr.add(1) as usize,
-                *base_ptr.add(2) as usize,
-            ]
-        }
+        let mesh = unsafe { &*self.mesh };
+        let base = (self.tri_index as usize) * 3;
+        [
+            mesh.vertex_indices[base] as usize,
+            mesh.vertex_indices[base + 1] as usize,
+            mesh.vertex_indices[base + 2] as usize,
+        ]
     }
 
-    #[inline(always)]
     fn get_points(&self) -> [Point3f; 3] {
+        let mesh = unsafe { &*self.mesh };
         let [v0, v1, v2] = self.get_vertex_indices();
-        unsafe {
-            [
-                *self.mesh.p.add(v0),
-                *self.mesh.p.add(v1),
-                *self.mesh.p.add(v2),
-            ]
-        }
+        [mesh.p[v0], mesh.p[v1], mesh.p[v2]]
     }
 
-    #[inline(always)]
-    fn get_uvs(&self) -> Option<[Point2f; 3]> {
-        if self.mesh.uv.is_null() {
-            return None;
-        }
-        let [v0, v1, v2] = self.get_vertex_indices();
-        unsafe {
-            Some([
-                *self.mesh.uv.add(v0),
-                *self.mesh.uv.add(v1),
-                *self.mesh.uv.add(v2),
-            ])
-        }
-    }
-
-    #[inline(always)]
-    fn get_tangents(&self) -> Option<[Vector3f; 3]> {
-        if self.mesh.s.is_null() {
-            return None;
-        }
-        let [v0, v1, v2] = self.get_vertex_indices();
-        unsafe {
-            Some([
-                *self.mesh.s.add(v0),
-                *self.mesh.s.add(v1),
-                *self.mesh.s.add(v2),
-            ])
-        }
-    }
-
-    #[inline(always)]
     fn get_shading_normals(&self) -> Option<[Normal3f; 3]> {
-        if self.mesh.n.is_null() {
+        let mesh = unsafe { &*self.mesh };
+        if mesh.n.is_empty() {
             return None;
         }
         let [v0, v1, v2] = self.get_vertex_indices();
-        unsafe {
-            Some([
-                *self.mesh.n.add(v0),
-                *self.mesh.n.add(v1),
-                *self.mesh.n.add(v2),
-            ])
-        }
+        Some([mesh.n[v0], mesh.n[v1], mesh.n[v2]])
     }
 
-    pub fn new(mesh: Ptr<DeviceTriangleMesh>, tri_index: i32) -> Self {
+    fn get_tangents(&self) -> Option<[Vector3f; 3]> {
+        let mesh = unsafe { &*self.mesh };
+        if mesh.s.is_empty() {
+            return None;
+        }
+        let [v0, v1, v2] = self.get_vertex_indices();
+        Some([mesh.s[v0], mesh.s[v1], mesh.s[v2]])
+    }
+
+    pub fn new(mesh: Ptr<TriangleMesh>, tri_index: i32) -> Self {
         Self { mesh, tri_index }
     }
 
-    pub fn get_mesh(&self) -> Ptr<DeviceTriangleMesh> {
+    pub fn get_mesh(&self) -> Ptr<TriangleMesh> {
         self.mesh
     }
 
diff --git a/shared/src/utils/containers.rs b/shared/src/utils/containers.rs
index 102e877..723a9f7 100644
--- a/shared/src/utils/containers.rs
+++ b/shared/src/utils/containers.rs
@@ -1,9 +1,8 @@
 use crate::core::geometry::{
     Bounds2i, Bounds3f, Bounds3i, Point2i, Point3f, Point3i, Vector2i, Vector3f, Vector3i,
 };
-use crate::core::pbrt::Float;
-use crate::utils::Ptr;
 use crate::utils::math::lerp;
+use crate::{gvec, gvec_from_slice, Float, GVec};
 use core::ops::{Add, Index, IndexMut, Mul, Sub};
 
 pub trait Interpolatable:
@@ -16,107 +15,88 @@ impl<T> Interpolatable for T where
 {
 }
 
-#[repr(C)]
-#[derive(Debug, Clone, Copy)]
-pub struct DeviceArray2D<T> {
-    pub values: *mut T,
-    pub extent: Bounds2i,
-    pub stride: i32,
+#[derive(Debug, Clone)]
+pub struct Array2D<T> {
+    extent: Bounds2i,
+    values: GVec<T>,
 }
 
-unsafe impl<T: Send> Send for DeviceArray2D<T> {}
-unsafe impl<T: Sync> Sync for DeviceArray2D<T> {}
-
-impl<T> DeviceArray2D<T> {
-    #[inline]
-    pub fn x_size(&self) -> u32 {
-        (self.extent.p_max.x() - self.extent.p_min.x()) as u32
+impl<T> Array2D<T> {
+    pub fn extent(&self) -> Bounds2i {
+        self.extent
     }
-
-    #[inline]
-    pub fn y_size(&self) -> u32 {
-        (self.extent.p_max.y() - self.extent.p_min.y()) as u32
+    pub fn stride(&self) -> i32 {
+        self.extent.p_max.x() - self.extent.p_min.x()
     }
-
-    #[inline]
-    pub fn size(&self) -> u32 {
-        self.extent.area() as u32
+    pub fn x_size(&self) -> usize {
+        self.stride() as usize
     }
-
-    #[inline(always)]
-    fn offset(&self, p: Point2i) -> isize {
-        let ox = p.x() - self.extent.p_min.x();
-        let oy = p.y() - self.extent.p_min.y();
-        (ox + oy * self.stride) as isize
+    pub fn y_size(&self) -> usize {
+        (self.extent.p_max.y() - self.extent.p_min.y()) as usize
     }
-
-    #[inline]
-    pub fn index(&self, x: i32, y: i32) -> u32 {
-        let nx = x - self.extent.p_min.x();
-        let ny = y - self.extent.p_min.y();
-        nx as u32 + self.x_size() * ny as u32
-    }
-
-    #[inline(always)]
-    pub fn get(&self, p: Point2i) -> &T {
-        unsafe { &*self.values.offset(self.offset(p)) }
-    }
-
-    #[inline(always)]
-    pub fn get_mut(&mut self, p: Point2i) -> &mut T {
-        unsafe { &mut *self.values.offset(self.offset(p)) }
-    }
-
-    #[inline]
-    pub fn get_linear(&self, index: usize) -> &T {
-        unsafe { &*self.values.add(index) }
-    }
-
-    #[inline]
-    pub fn get_linear_mut(&mut self, index: usize) -> &mut T {
-        unsafe { &mut *self.values.add(index) }
-    }
-
     pub fn as_slice(&self) -> &[T] {
-        unsafe { core::slice::from_raw_parts(self.values, self.size() as usize) }
+        &self.values
     }
-
     pub fn as_mut_slice(&mut self) -> &mut [T] {
-        unsafe { core::slice::from_raw_parts_mut(self.values, self.size() as usize) }
+        &mut self.values
+    }
+    pub fn as_ptr(&self) -> *const T {
+        self.values.as_ptr()
+    }
+    pub fn as_mut_ptr(&mut self) -> *mut T {
+        self.values.as_mut_ptr()
+    }
+    pub fn len(&self) -> usize {
+        self.values.len()
     }
 }
 
-impl<T> Index<Point2i> for DeviceArray2D<T> {
+unsafe impl<T: Send> Send for Array2D<T> {}
+unsafe impl<T: Sync> Sync for Array2D<T> {}
+
+impl<T: Default + Clone> Array2D<T> {
+    pub fn new(extent: Bounds2i) -> Self {
+        let n = extent.area() as usize;
+        let mut values = gvec();
+        values.resize(n, T::default());
+        Self { extent, values }
+    }
+
+    pub fn new_dims(nx: i32, ny: i32) -> Self {
+        Self::new(Bounds2i::from_points(
+            Point2i::new(0, 0),
+            Point2i::new(nx, ny),
+        ))
+    }
+
+    pub fn new_filled(nx: i32, ny: i32, val: T) -> Self {
+        let extent = Bounds2i::from_points(Point2i::new(0, 0), Point2i::new(nx, ny));
+        let n = (nx * ny) as usize;
+        let mut values = gvec();
+        values.resize(n, val);
+        Self { extent, values }
+    }
+}
+
+impl<T> Index<(i32, i32)> for Array2D<T> {
     type Output = T;
-    #[inline(always)]
-    fn index(&self, p: Point2i) -> &Self::Output {
-        self.get(p)
+    fn index(&self, (x, y): (i32, i32)) -> &T {
+        let offset = (y - self.extent.p_min.y()) * self.stride() + (x - self.extent.p_min.x());
+        &self.values[offset as usize]
     }
 }
 
-impl<T> IndexMut<Point2i> for DeviceArray2D<T> {
-    fn index_mut(&mut self, p: Point2i) -> &mut Self::Output {
-        self.get_mut(p)
-    }
-}
-
-impl<T> Index<(i32, i32)> for DeviceArray2D<T> {
-    type Output = T;
-    fn index(&self, (x, y): (i32, i32)) -> &Self::Output {
-        &self[Point2i::new(x, y)]
-    }
-}
-
-impl<T> IndexMut<(i32, i32)> for DeviceArray2D<T> {
-    fn index_mut(&mut self, (x, y): (i32, i32)) -> &mut Self::Output {
-        &mut self[Point2i::new(x, y)]
+impl<T> IndexMut<(i32, i32)> for Array2D<T> {
+    fn index_mut(&mut self, (x, y): (i32, i32)) -> &mut T {
+        let offset = (y - self.extent.p_min.y()) * self.stride() + (x - self.extent.p_min.x());
+        &mut self.values[offset as usize]
     }
 }
 
 #[repr(C)]
 #[derive(Debug, Clone, Copy)]
 pub struct SampledGrid<T> {
-    pub values: Ptr<T>,
+    pub values: GVec<T>,
     pub values_len: u32,
     pub nx: i32,
     pub ny: i32,
@@ -127,11 +107,10 @@ unsafe impl<T: Sync> Sync for SampledGrid<T> {}
 unsafe impl<T: Send> Send for SampledGrid<T> {}
 
 impl<T> SampledGrid<T> {
-    #[cfg(not(target_os = "cuda"))]
     pub fn new(slice: &[T], nx: i32, ny: i32, nz: i32) -> Self {
         assert_eq!(slice.len(), (nx * ny * nz) as usize);
         Self {
-            values: Ptr::from(slice),
+            values: gvec_from_slice(slice),
             values_len: (nx * ny * nz) as u32,
             nx,
             ny,
@@ -141,7 +120,7 @@ impl<T> SampledGrid<T> {
 
     pub fn empty() -> Self {
         Self {
-            values: Ptr::null(),
+            values: gvec(),
             values_len: 0,
             nx: 0,
             ny: 0,
diff --git a/shared/src/utils/math.rs b/shared/src/utils/math.rs
index 8685554..426ec27 100644
--- a/shared/src/utils/math.rs
+++ b/shared/src/utils/math.rs
@@ -137,7 +137,12 @@ pub fn windowed_sinc(x: Float, radius: Float, tau: Float) -> Float {
     if x.abs() > radius {
         return 0.;
     }
-    sinc(x) * sinc(x / tau)
+
+    if x < 1e-5 {
+        1.0
+    } else {
+        sinc(x) * sinc(x / tau)
+    }
 }
 
 #[inline]
diff --git a/shared/src/utils/mesh.rs b/shared/src/utils/mesh.rs
deleted file mode 100644
index e091c93..0000000
--- a/shared/src/utils/mesh.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-use crate::Float;
-use crate::core::geometry::{Normal3f, Point2f, Point3f, Vector3f};
-use crate::utils::Ptr;
-use crate::utils::Transform;
-use crate::utils::sampling::DevicePiecewiseConstant2D;
-
-#[repr(C)]
-#[derive(Debug, Copy, Clone)]
-pub struct DeviceTriangleMesh {
-    pub p: Ptr<Point3f>,
-    pub n: Ptr<Normal3f>,
-    pub s: Ptr<Vector3f>,
-    pub uv: Ptr<Point2f>,
-    pub vertex_indices: Ptr<i32>,
-    pub face_indices: Ptr<i32>,
-    pub n_triangles: u32,
-    pub n_vertices: u32,
-    pub reverse_orientation: bool,
-    pub transform_swaps_handedness: bool,
-}
-
-#[repr(C)]
-#[derive(Debug, Clone, Copy)]
-pub struct DeviceBilinearPatchMesh {
-    pub image_distribution: Ptr<DevicePiecewiseConstant2D>,
-    pub p: Ptr<Point3f>,
-    pub n: Ptr<Normal3f>,
-    pub uv: Ptr<Point2f>,
-    pub vertex_indices: Ptr<i32>,
-    pub n_patches: u32,
-    pub n_vertices: u32,
-    pub reverse_orientation: bool,
-    pub transform_swaps_handedness: bool,
-}
-
-unsafe impl Send for DeviceTriangleMesh {}
-unsafe impl Sync for DeviceTriangleMesh {}
-unsafe impl Send for DeviceBilinearPatchMesh {}
-unsafe impl Sync for DeviceBilinearPatchMesh {}
diff --git a/shared/src/utils/mod.rs b/shared/src/utils/mod.rs
index a18ea95..b160d1d 100644
--- a/shared/src/utils/mod.rs
+++ b/shared/src/utils/mod.rs
@@ -3,7 +3,6 @@ pub mod containers;
 pub mod hash;
 pub mod interval;
 pub mod math;
-pub mod mesh;
 pub mod noise;
 pub mod options;
 pub mod ptr;
@@ -18,15 +17,6 @@ pub use options::PBRTOptions;
 pub use ptr::Ptr;
 pub use transform::{AnimatedTransform, Transform, TransformGeneric};
 
-use proc_macro::TokenStream;
-use proc_macro2::TokenStream as TokenStream2;
-use quote::{format_ident, quote};
-use syn::{
-    parse_macro_input, Attribute, Data, DeriveInput, Expr, Fields, GenericArgument, Ident, Lit,
-    PathArguments, Type, Variant,
-};
-
-
 use crate::Float;
 use core::sync::atomic::{AtomicU32, Ordering};
 
@@ -138,514 +128,3 @@ pub fn gpu_array_from_fn<T, const N: usize>(mut f: impl FnMut(usize) -> T) -> [T
     }
 }
 
-/// # Enum variant attributes
-///
-/// | Attribute | Effect |
-/// |-----------|--------|
-/// | *(none)* | Inner type has `DeviceRepr`; auto-call `upload_value` |
-/// | `#[device(clone)]` | Same type on both sides, just clone |
-/// | `#[device(custom = "method")]` | You provide `fn method(inner: &T, arena) -> DeviceT` |
-/// | `#[device(variant_type = "T")]` | Override the device-side variant's inner type |
-///
-/// # Container attribute
-///
-/// `#[device(name = "DeviceFoo")]` — override the generated type name (default: `Device{Name}`).
-#[proc_macro_derive(Device, attributes(device))]
-pub fn derive_device(input: TokenStream) -> TokenStream {
-    let input = parse_macro_input!(input as DeriveInput);
-    match derive_impl(input) {
-        Ok(tokens) => tokens.into(),
-        Err(e) => e.to_compile_error().into(),
-    }
-}
-
-fn derive_impl(input: DeriveInput) -> syn::Result<TokenStream2> {
-    match &input.data {
-        Data::Struct(_) => derive_struct(input),
-        Data::Enum(_) => derive_enum(input),
-        Data::Union(_) => Err(syn::Error::new_spanned(
-            &input.ident,
-            "Device derive does not support unions",
-        )),
-    }
-}
-
-// Struct derivation
-
-fn derive_struct(input: DeriveInput) -> syn::Result<TokenStream2> {
-    let host_name = &input.ident;
-    let vis = &input.vis;
-    let device_name = get_device_name(&input.attrs, host_name)?;
-
-    let fields = match &input.data {
-        Data::Struct(s) => match &s.fields {
-            Fields::Named(named) => &named.named,
-            _ => {
-                return Err(syn::Error::new_spanned(
-                    host_name,
-                    "Device derive only supports structs with named fields",
-                ))
-            }
-        },
-        _ => unreachable!(),
-    };
-
-    let mut device_fields = Vec::new();
-    let mut upload_stmts = Vec::new();
-    let mut device_field_inits = Vec::new();
-    let mut spread_expr: Option<Expr> = None;
-
-    for field in fields {
-        let field_name = field.ident.as_ref().unwrap();
-        let attrs = parse_field_attrs(&field.attrs)?;
-
-        if attrs.skip {
-            continue;
-        }
-
-        if let Some(ref expr_str) = attrs.spread {
-            spread_expr = Some(syn::parse_str(expr_str).map_err(|e| {
-                syn::Error::new_spanned(field, format!("invalid device(spread): {}", e))
-            })?);
-            continue;
-        }
-
-        if let Some(expr_str) = &attrs.expr {
-            let expr: Expr = syn::parse_str(expr_str).map_err(|e| {
-                syn::Error::new_spanned(field, format!("invalid device(expr): {}", e))
-            })?;
-            let ty = &field.ty;
-            device_fields.push(quote! { pub #field_name: #ty });
-            upload_stmts.push(quote! { let #field_name = #expr; });
-            device_field_inits.push(quote! { #field_name });
-            continue;
-        }
-
-        match classify_type(&field.ty) {
-            FieldClass::VecCopy(inner_ty) => {
-                let len_name = format_ident!("{}_len", field_name);
-                device_fields.push(quote! { pub #field_name: Ptr<#inner_ty> });
-                device_fields.push(quote! { pub #len_name: usize });
-                upload_stmts.push(quote! {
-                    let (#field_name, #len_name) = arena.alloc_slice(&self.#field_name);
-                });
-                device_field_inits.push(quote! { #field_name });
-                device_field_inits.push(quote! { #len_name });
-            }
-            FieldClass::VecUploadable(inner_ty) => {
-                let len_name = format_ident!("{}_len", field_name);
-                device_fields.push(quote! {
-                    pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
-                });
-                device_fields.push(quote! { pub #len_name: usize });
-                upload_stmts.push(quote! {
-                    let __up: Vec<<#inner_ty as DeviceRepr>::Target> = self.#field_name
-                        .iter()
-                        .map(|item| DeviceRepr::upload_value(item, arena))
-                        .collect();
-                    let (#field_name, #len_name) = arena.alloc_slice(&__up);
-                });
-                device_field_inits.push(quote! { #field_name });
-                device_field_inits.push(quote! { #len_name });
-            }
-            FieldClass::Option(inner_ty) => {
-                device_fields.push(quote! {
-                    pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
-                });
-                upload_stmts.push(quote! {
-                    let #field_name = match &self.#field_name {
-                        Some(val) => DeviceRepr::upload(val, arena),
-                        None => Ptr::null(),
-                    };
-                });
-                device_field_inits.push(quote! { #field_name });
-            }
-            FieldClass::Arc(inner_ty) => {
-                if attrs.flatten {
-                    device_fields.push(quote! {
-                        pub #field_name: <#inner_ty as DeviceRepr>::Target
-                    });
-                    upload_stmts.push(quote! {
-                        let #field_name = DeviceRepr::upload_value(&*self.#field_name, arena);
-                    });
-                } else {
-                    device_fields.push(quote! {
-                        pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
-                    });
-                    upload_stmts.push(quote! {
-                        let #field_name = DeviceRepr::upload(&*self.#field_name, arena);
-                    });
-                }
-                device_field_inits.push(quote! { #field_name });
-            }
-            FieldClass::Plain => {
-                let ty = &field.ty;
-                if attrs.copy_upload {
-                    device_fields.push(quote! { pub #field_name: #ty });
-                    upload_stmts.push(quote! {
-                        let #field_name = self.#field_name.clone();
-                    });
-                } else if attrs.flatten {
-                    device_fields.push(quote! {
-                        pub #field_name: <#ty as DeviceRepr>::Target
-                    });
-                    upload_stmts.push(quote! {
-                        let #field_name = DeviceRepr::upload_value(&self.#field_name, arena);
-                    });
-                } else if attrs.upload {
-                    device_fields.push(quote! {
-                        pub #field_name: Ptr<<#ty as DeviceRepr>::Target>
-                    });
-                    upload_stmts.push(quote! {
-                        let #field_name = DeviceRepr::upload(&self.#field_name, arena);
-                    });
-                } else {
-                    device_fields.push(quote! { pub #field_name: #ty });
-                    upload_stmts.push(quote! {
-                        let #field_name = self.#field_name;
-                    });
-                }
-                device_field_inits.push(quote! { #field_name });
-            }
-        }
-    }
-
-    let constructor = if let Some(spread) = spread_expr {
-        quote! {
-            #device_name {
-                #(#device_field_inits,)*
-                ..#spread
-            }
-        }
-    } else {
-        quote! {
-            #device_name {
-                #(#device_field_inits,)*
-            }
-        }
-    };
-
-    Ok(quote! {
-        #[repr(C)]
-        #[derive(Debug, Copy, Clone)]
-        #vis struct #device_name {
-            #(#device_fields,)*
-        }
-
-        unsafe impl Send for #device_name {}
-        unsafe impl Sync for #device_name {}
-
-        impl DeviceRepr for #host_name {
-            type Target = #device_name;
-
-            fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
-                #(#upload_stmts)*
-                #constructor
-            }
-        }
-    })
-}
-
-// Enum derivation
-fn derive_enum(input: DeriveInput) -> syn::Result<TokenStream2> {
-    let host_name = &input.ident;
-    let vis = &input.vis;
-    let device_name = get_device_name(&input.attrs, host_name)?;
-
-    let variants = match &input.data {
-        Data::Enum(e) => &e.variants,
-        _ => unreachable!(),
-    };
-
-    let mut device_variants = Vec::new();
-    let mut match_arms = Vec::new();
-
-    for variant in variants {
-        let var_name = &variant.ident;
-        let var_attrs = parse_variant_attrs(&variant.attrs)?;
-        let inner_ty = get_variant_inner_type(variant)?;
-
-        // Determine the device-side inner type for this variant
-        let device_inner: Type = if let Some(ref ty_str) = var_attrs.variant_type {
-            syn::parse_str(ty_str).map_err(|e| {
-                syn::Error::new_spanned(variant, format!("invalid variant_type: {}", e))
-            })?
-        } else if var_attrs.clone_variant {
-            // clone: same type on both sides
-            inner_ty.clone()
-        } else {
-            // auto-upload: use DeviceRepr::Target
-            syn::parse_str(&format!("<{} as DeviceRepr>::Target", quote!(#inner_ty))).map_err(
-                |e| {
-                    syn::Error::new_spanned(variant, format!("cannot construct Target type: {}", e))
-                },
-            )?
-        };
-
-        device_variants.push(quote! { #var_name(#device_inner) });
-
-        if var_attrs.clone_variant {
-            match_arms.push(quote! {
-                #host_name::#var_name(inner) => #device_name::#var_name(inner.clone())
-            });
-        } else if let Some(ref method) = var_attrs.custom {
-            let method_ident = format_ident!("{}", method);
-            match_arms.push(quote! {
-                #host_name::#var_name(inner) => {
-                    #device_name::#var_name(Self::#method_ident(inner, arena))
-                }
-            });
-        } else {
-            // Default: inner implements DeviceRepr
-            match_arms.push(quote! {
-                #host_name::#var_name(inner) => {
-                    #device_name::#var_name(DeviceRepr::upload_value(inner, arena))
-                }
-            });
-        }
-    }
-
-    Ok(quote! {
-        #[repr(C)]
-        #[derive(Debug, Copy, Clone)]
-        #vis enum #device_name {
-            #(#device_variants,)*
-        }
-
-        unsafe impl Send for #device_name {}
-        unsafe impl Sync for #device_name {}
-
-        impl DeviceRepr for #host_name {
-            type Target = #device_name;
-
-            fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
-                match self {
-                    #(#match_arms,)*
-                }
-            }
-        }
-    })
-}
-
-fn get_variant_inner_type(variant: &Variant) -> syn::Result<Type> {
-    match &variant.fields {
-        Fields::Unnamed(fields) if fields.unnamed.len() == 1 => {
-            Ok(fields.unnamed.first().unwrap().ty.clone())
-        }
-        Fields::Unit => Err(syn::Error::new_spanned(
-            variant,
-            "Device derive: enum variants must have exactly one field, e.g. Variant(Type)",
-        )),
-        _ => Err(syn::Error::new_spanned(
-            variant,
-            "Device derive: only single-field tuple variants supported, e.g. Variant(Type)",
-        )),
-    }
-}
-
-// Attribute parsing for variants
-struct VariantAttrs {
-    clone_variant: bool,
-    custom: Option<String>,
-    variant_type: Option<String>,
-}
-
-fn parse_variant_attrs(attrs: &[Attribute]) -> syn::Result<VariantAttrs> {
-    let mut result = VariantAttrs {
-        clone_variant: false,
-        custom: None,
-        variant_type: None,
-    };
-
-    for attr in attrs {
-        if !attr.path().is_ident("device") {
-            continue;
-        }
-        attr.parse_nested_meta(|meta| {
-            if meta.path.is_ident("clone") {
-                result.clone_variant = true;
-                Ok(())
-            } else if meta.path.is_ident("custom") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    result.custom = Some(s.value());
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else if meta.path.is_ident("variant_type") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    result.variant_type = Some(s.value());
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else {
-                Err(meta.error("unknown device variant attribute"))
-            }
-        })?;
-    }
-
-    Ok(result)
-}
-
-// Attribute parsing for fields
-struct FieldAttrs {
-    skip: bool,
-    expr: Option<String>,
-    copy_upload: bool,
-    flatten: bool,
-    upload: bool,
-    spread: Option<String>,
-}
-
-fn parse_field_attrs(attrs: &[Attribute]) -> syn::Result<FieldAttrs> {
-    let mut result = FieldAttrs {
-        skip: false,
-        expr: None,
-        copy_upload: false,
-        flatten: false,
-        upload: false,
-        spread: None,
-    };
-
-    for attr in attrs {
-        if !attr.path().is_ident("device") {
-            continue;
-        }
-        attr.parse_nested_meta(|meta| {
-            if meta.path.is_ident("skip") {
-                result.skip = true;
-                Ok(())
-            } else if meta.path.is_ident("expr") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    result.expr = Some(s.value());
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else if meta.path.is_ident("copy_upload") {
-                result.copy_upload = true;
-                Ok(())
-            } else if meta.path.is_ident("flatten") {
-                result.flatten = true;
-                Ok(())
-            } else if meta.path.is_ident("upload") {
-                result.upload = true;
-                Ok(())
-            } else if meta.path.is_ident("spread") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    result.spread = Some(s.value());
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else {
-                Err(meta.error("unknown device attribute"))
-            }
-        })?;
-    }
-
-    Ok(result)
-}
-
-// Container-level name attribute
-fn get_device_name(attrs: &[Attribute], host_name: &Ident) -> syn::Result<Ident> {
-    for attr in attrs {
-        if !attr.path().is_ident("device") {
-            continue;
-        }
-        let mut name = None;
-        attr.parse_nested_meta(|meta| {
-            if meta.path.is_ident("name") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    name = Some(format_ident!("{}", s.value()));
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else {
-                Ok(())
-            }
-        })?;
-        if let Some(n) = name {
-            return Ok(n);
-        }
-    }
-    Ok(format_ident!("Device{}", host_name))
-}
-
-// Type classification
-enum FieldClass {
-    VecCopy(Type),
-    VecUploadable(Type),
-    Option(Type),
-    Arc(Type),
-    Plain,
-}
-
-fn classify_type(ty: &Type) -> FieldClass {
-    if let Some(inner) = extract_generic_arg(ty, "Vec") {
-        if is_copy_primitive(&inner) {
-            FieldClass::VecCopy(inner)
-        } else {
-            FieldClass::VecUploadable(inner)
-        }
-    } else if let Some(inner) = extract_generic_arg(ty, "Option") {
-        FieldClass::Option(inner)
-    } else if let Some(inner) = extract_generic_arg(ty, "Arc") {
-        FieldClass::Arc(inner)
-    } else {
-        FieldClass::Plain
-    }
-}
-
-fn extract_generic_arg(ty: &Type, wrapper: &str) -> Option<Type> {
-    if let Type::Path(type_path) = ty {
-        let seg = type_path.path.segments.last()?;
-        if seg.ident != wrapper {
-            return None;
-        }
-        if let PathArguments::AngleBracketed(args) = &seg.arguments {
-            if let Some(GenericArgument::Type(inner)) = args.args.first() {
-                return Some(inner.clone());
-            }
-        }
-    }
-    None
-}
-
-fn is_copy_primitive(ty: &Type) -> bool {
-    if let Type::Path(type_path) = ty {
-        if let Some(seg) = type_path.path.segments.last() {
-            let name = seg.ident.to_string();
-            return matches!(
-                name.as_str(),
-                "f32"
-                    | "f64"
-                    | "u8"
-                    | "u16"
-                    | "u32"
-                    | "u64"
-                    | "i8"
-                    | "i16"
-                    | "i32"
-                    | "i64"
-                    | "usize"
-                    | "isize"
-                    | "bool"
-                    | "Float"
-            );
-        }
-    }
-    false
-}
diff --git a/shared/src/utils/sampling.rs b/shared/src/utils/sampling.rs
index 5a8bd2a..db63650 100644
--- a/shared/src/utils/sampling.rs
+++ b/shared/src/utils/sampling.rs
@@ -1,7 +1,7 @@
 use crate::core::geometry::{
     Bounds2f, Frame, Point2f, Point2i, Point3f, Vector2f, Vector2i, Vector3f, VectorLike,
 };
-use crate::utils::containers::DeviceArray2D;
+use crate::{GVec, gvec_with_capacity, gvec_from_slice, Array2D};
 use crate::utils::find_interval;
 use crate::utils::math::{
     catmull_rom_weights, clamp, difference_of_products, evaluate_polynomial, lerp, logistic,
@@ -14,7 +14,7 @@ use num_traits::Float as NumFloat;
 use num_traits::Num;
 
 #[cfg(feature = "cpu_debug")]
-use crate::{RARE_EVENT_CONDITION_MET, RARE_EVENT_TOTAL_CALLS, check_rare};
+use crate::{check_rare, RARE_EVENT_CONDITION_MET, RARE_EVENT_TOTAL_CALLS};
 
 pub fn linear_pdf<T>(x: T, a: T, b: T) -> T
 where
@@ -703,128 +703,177 @@ pub struct PLSample {
 }
 
 #[repr(C)]
-#[derive(Debug, Copy, Clone)]
-pub struct DevicePiecewiseConstant1D {
-    pub func: Ptr<Float>,
-    pub cdf: Ptr<Float>,
+#[derive(Debug, Clone)]
+pub struct PiecewiseConstant1D {
+    pub func: GVec<Float>,
+    pub cdf: GVec<Float>,
     pub min: Float,
     pub max: Float,
-    pub n: u32,
     pub func_integral: Float,
 }
 
-unsafe impl Send for DevicePiecewiseConstant1D {}
-unsafe impl Sync for DevicePiecewiseConstant1D {}
+unsafe impl Send for PiecewiseConstant1D {}
+unsafe impl Sync for PiecewiseConstant1D {}
 
-impl DevicePiecewiseConstant1D {
+impl PiecewiseConstant1D {
+    pub fn new(f: &[Float]) -> Self {
+        Self::new_with_bounds(f, 0.0, 1.0)
+    }
+
+    pub fn new_with_bounds(f: &[Float], min: Float, max: Float) -> Self {
+        let n = f.len();
+        let mut cdf = gvec_with_capacity(n + 1);
+        cdf.push(0.0);
+
+        let delta = (max - min) / n as Float;
+        for i in 0..n {
+            cdf.push(cdf[i] + f[i] * delta);
+        }
+
+        let func_integral = cdf[n];
+        if func_integral > 0.0 {
+            for c in &mut cdf {
+                *c /= func_integral;
+            }
+        }
+
+        Self {
+            func: gvec_from_slice(f),
+            cdf,
+            min,
+            max,
+            func_integral,
+        }
+    }
+
+    pub fn from_func<F>(f: F, min: Float, max: Float, n: usize) -> Self
+    where
+        F: Fn(Float) -> Float,
+    {
+        let delta = (max - min) / n as Float;
+        let values: Vec<Float> = (0..n)
+            .map(|i| f(min + (i as Float + 0.5) * delta))
+            .collect();
+        Self::new_with_bounds(&values, min, max)
+    }
+
+    pub fn n(&self) -> usize {
+        self.func.len()
+    }
+    pub fn func(&self) -> &[Float] {
+        &self.func
+    }
+    pub fn cdf(&self) -> &[Float] {
+        &self.cdf
+    }
     pub fn integral(&self) -> Float {
         self.func_integral
     }
 
-    pub fn size(&self) -> u32 {
-        self.n
-    }
-
-    fn find_interval(&self, u: Float) -> usize {
-        let mut size = self.n as usize;
-        let mut first = 0;
-
+    pub fn find_interval(&self, u: Float) -> usize {
+        let n = self.func.len();
+        let mut size = n;
+        let mut first = 0usize;
         while size > 0 {
             let half = size >> 1;
             let middle = first + half;
-
-            let cdf_val = unsafe { *self.cdf.add(middle) };
-
-            if cdf_val <= u {
+            if self.cdf[middle] <= u {
                 first = middle + 1;
                 size -= half + 1;
             } else {
                 size = half;
             }
         }
-
-        (first - 1).clamp(0, self.n as usize - 1)
+        first.saturating_sub(1).min(n - 1)
     }
 
     pub fn sample(&self, u: Float) -> (Float, Float, usize) {
-        // Find offset via binary search on CDF
         let offset = self.find_interval(u);
-
-        let cdf_offset = unsafe { *self.cdf.add(offset) };
-        let cdf_next = unsafe { *self.cdf.add(offset + 1) };
-
+        let cdf_offset = self.cdf[offset];
+        let cdf_next = self.cdf[offset + 1];
         let du = if cdf_next - cdf_offset > 0.0 {
             (u - cdf_offset) / (cdf_next - cdf_offset)
         } else {
             0.0
         };
-
-        let delta = (self.max - self.min) / self.n as Float;
+        let n = self.func.len();
+        let delta = (self.max - self.min) / n as Float;
         let x = self.min + (offset as Float + du) * delta;
-
         let pdf = if self.func_integral > 0.0 {
-            (unsafe { *self.func.add(offset) }) / self.func_integral
+            self.func[offset] / self.func_integral
         } else {
             0.0
         };
-
         (x, pdf, offset)
     }
 }
 
 #[repr(C)]
-#[derive(Debug, Copy, Clone)]
-pub struct DevicePiecewiseConstant2D {
-    pub conditionals: Ptr<DevicePiecewiseConstant1D>, // Array of n_v conditionals
-    pub marginal: DevicePiecewiseConstant1D,
+#[derive(Debug, Clone)]
+pub struct PiecewiseConstant2D {
+    pub conditionals: GVec<PiecewiseConstant1D>,
+    pub marginal: PiecewiseConstant1D,
     pub n_u: u32,
     pub n_v: u32,
 }
 
-unsafe impl Send for DevicePiecewiseConstant2D {}
-unsafe impl Sync for DevicePiecewiseConstant2D {}
+impl PiecewiseConstant2D {
+    pub fn new(data: &Array2D<Float>) -> Self {
+        Self::new_with_bounds(data, Bounds2f::unit())
+    }
 
-impl DevicePiecewiseConstant2D {
-    // pub fn resolution(&self) -> Point2i {
-    //     Point2i::new(
-    //         self.p_conditional_v[0u32].size() as i32,
-    //         self.p_conditional_v[1u32].size() as i32,
-    //     )
-    // }
+    pub fn new_with_bounds(data: &Array2D<Float>, domain: Bounds2f) -> Self {
+        Self::from_slice(data.as_slice(), data.x_size(), data.y_size(), domain)
+    }
 
-    pub fn integral(&self) -> f32 {
+    pub fn from_slice(data: &[Float], n_u: usize, n_v: usize, domain: Bounds2f) -> Self {
+        assert_eq!(data.len(), n_u * n_v);
+
+        let mut conditionals = gvec_with_capacity(n_v);
+        let mut marginal_func = GVec::with_capacity(n_v);
+
+        for v in 0..n_v {
+            let row = &data[v * n_u..(v + 1) * n_u];
+            let conditional = PiecewiseConstant1D::new_with_bounds(
+                row, domain.p_min.x(), domain.p_max.x(),
+            );
+            marginal_func.push(conditional.integral());
+            conditionals.push(conditional);
+        }
+
+        let marginal = PiecewiseConstant1D::new_with_bounds(
+            &marginal_func, domain.p_min.y(), domain.p_max.y(),
+        );
+
+        Self {
+            conditionals,
+            marginal,
+            n_u: n_u as u32,
+            n_v: n_v as u32,
+        }
+    }
+
+    pub fn integral(&self) -> Float {
         self.marginal.integral()
     }
 
-    pub fn sample(&self, u: Point2f) -> (Point2f, f32, Point2i) {
+    pub fn sample(&self, u: Point2f) -> (Point2f, Float, Point2i) {
         let (d1, pdf1, off_y) = self.marginal.sample(u.y());
-        let (d0, pdf0, off_x) = (unsafe { self.conditionals.add(off_y) }).sample(u.x());
+        let (d0, pdf0, off_x) = self.conditionals[off_y].sample(u.x());
         let pdf = pdf0 * pdf1;
-        let offset = Point2i::new(off_x as i32, off_y as i32);
-        (Point2f::new(d0, d1), pdf, offset)
-    }
-
-    pub fn pdf(&self, p: Point2f) -> Float {
-        // Find which row
-        // let delta_v = 1.0 / self.n_v as Float;
-        let v_offset = ((p.y() * self.n_v as Float) as usize).min(self.n_v as usize - 1);
-
-        let conditional = unsafe { &*self.conditionals.add(v_offset) };
-
-        // Find which column
-        // let delta_u = 1.0 / self.n_u as Float;
-        let u_offset = ((p.x() * self.n_u as Float) as usize).min(self.n_u as usize - 1);
-
-        let func_val = unsafe { *conditional.func.add(u_offset) };
-
-        func_val / self.marginal.func_integral
+        (
+            Point2f::new(d0, d1),
+            pdf,
+            Point2i::new(off_x as i32, off_y as i32),
+        )
     }
 }
 
+
 #[repr(C)]
 #[derive(Debug, Copy, Clone)]
 pub struct DeviceSummedAreaTable {
-    pub sum: DeviceArray2D<f64>,
+    pub sum: Array2D<f64>,
 }
 
 impl DeviceSummedAreaTable {
@@ -877,7 +926,7 @@ impl DeviceSummedAreaTable {
 #[derive(Debug, Copy, Clone)]
 pub struct DeviceWindowedPiecewiseConstant2D {
     pub sat: DeviceSummedAreaTable,
-    pub func: DeviceArray2D<Float>,
+    pub func: Array2D<Float>,
 }
 
 impl DeviceWindowedPiecewiseConstant2D {
@@ -1058,10 +1107,10 @@ pub struct PiecewiseLinear2D<const N: usize> {
     pub inv_patch_size: Vector2f,
     pub param_size: [u32; N],
     pub param_strides: [u32; N],
-    pub param_values: [Ptr<Float>; N],
-    pub data: Ptr<Float>,
-    pub marginal_cdf: Ptr<Float>,
-    pub conditional_cdf: Ptr<Float>,
+    pub param_values: [GVec<Float>; N],
+    pub data: GVec<Float>,
+    pub marginal_cdf: GVec<Float>,
+    pub conditional_cdf: GVec<Float>,
 }
 
 impl<const N: usize> PiecewiseLinear2D<N> {
diff --git a/src/core/aggregates.rs b/src/core/aggregates.rs
index 3fda232..fc03238 100644
--- a/src/core/aggregates.rs
+++ b/src/core/aggregates.rs
@@ -1,5 +1,5 @@
 use rayon::prelude::*;
-use shared::core::aggregates::{{DeviceBVHAggregate, LinearBVHNode};
+use shared::core::aggregates::{DeviceBVHAggregate, LinearBVHNode};
 use shared::core::geometry::{Bounds3f, Point3f, Ray, Vector3f};
 use shared::core::primitive::{Primitive, PrimitiveTrait};
 use shared::core::shape::ShapeIntersection;
@@ -307,7 +307,7 @@ impl<P: PrimitiveTrait + Clone + Send + Sync> BVHAggregate<P> {
     pub fn build_hlbvh(
         bvh_primitives: &[BVHPrimitiveInfo],
         total_nodes: &AtomicUsize,
-        original_primitives: &[P],
+        _original_primitives: &[P],
         max_prims_in_node: usize,
     ) -> Box<BVHBuildNode> {
         let bounds = bvh_primitives
diff --git a/src/core/camera.rs b/src/core/camera.rs
index 8985d80..7263976 100644
--- a/src/core/camera.rs
+++ b/src/core/camera.rs
@@ -388,11 +388,11 @@ impl CameraFactory for Camera {
                     Arc::from(aperture_image.unwrap()),
                 );
 
-                // arena.alloc(camera);
+                arena.alloc(camera);
                 Ok(Camera::Realistic(camera.device()))
             }
             "spherical" => {
-                let full_res = film.full_resolution();
+                let _full_res = film.full_resolution();
                 let camera_params =
                     CameraBaseParameters::new(camera_transform, film, medium, params, loc)?;
                 let base = CameraBase::create(camera_params);
diff --git a/src/core/film.rs b/src/core/film.rs
index ce291da..b19ea73 100644
--- a/src/core/film.rs
+++ b/src/core/film.rs
@@ -1,26 +1,26 @@
-use crate::Arena;
 use crate::core::image::{Image, ImageChannelDesc, ImageChannelValues, ImageIO, ImageMetadata};
 use crate::films::*;
 use crate::spectra::data::get_named_spectrum;
 use crate::spectra::piecewise::PiecewiseLinearSpectrumBuffer;
-use anyhow::{Result, anyhow};
+use crate::Arena;
+use anyhow::{anyhow, Result};
 use rayon::iter::ParallelIterator;
 use rayon::prelude::IntoParallelIterator;
 use shared::core::camera::CameraTransform;
-use shared::core::color::{RGB, SRGB, XYZ, white_balance};
+use shared::core::color::{white_balance, RGB, SRGB, XYZ};
 use shared::core::film::{DevicePixelSensor, Film, FilmBase, GBufferFilm, RGBFilm, SpectralFilm};
 use shared::core::filter::{Filter, FilterTrait};
 use shared::core::geometry::{Bounds2f, Bounds2i, Point2f, Point2i};
 use shared::core::image::PixelFormat;
 use shared::core::spectrum::Spectrum;
-use shared::spectra::{RGBColorSpace, cie::SWATCHES_RAW};
-use shared::utils::math::{SquareMatrix, linear_least_squares};
+use shared::spectra::{cie::SWATCHES_RAW, RGBColorSpace};
+use shared::utils::math::{linear_least_squares, SquareMatrix};
 use shared::{Float, Ptr};
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, LazyLock};
 
 use crate::spectra::{
-    CIE_X_DATA, CIE_Y_DATA, CIE_Z_DATA, DenselySampledSpectrumBuffer, get_spectra_context,
+    get_spectra_context, DenselySampledSpectrumBuffer, CIE_X_DATA, CIE_Y_DATA, CIE_Z_DATA,
 };
 use crate::utils::{FileLoc, ParameterDictionary};
 
@@ -363,7 +363,7 @@ pub trait FilmTrait: Sync {
             .collect();
 
         let mut image = Image::new(format, resolution, channel_names, SRGB.into());
-        let rgb_desc = ImageChannelDesc::new(&[0, 1, 2]);
+        let _rgb_desc = ImageChannelDesc::new(&[0, 1, 2]);
 
         for (iy, row_data) in processed_rows.into_iter().enumerate() {
             for (ix, rgb_chunk) in row_data.chunks_exact(3).enumerate() {
@@ -381,10 +381,6 @@ pub trait FilmTrait: Sync {
             );
         }
 
-        // self.base().pixel_bounds = pixel_bounds;
-        // self.base().full_resolution = resolution;
-        // self.colorspace = colorspace;
-
         image
     }
 }
diff --git a/src/core/filter.rs b/src/core/filter.rs
index b8874d2..65ed690 100644
--- a/src/core/filter.rs
+++ b/src/core/filter.rs
@@ -1,20 +1,28 @@
-use crate::filters::*;
-use crate::utils::containers::Array2D;
 use crate::utils::sampling::PiecewiseConstant2D;
-use crate::utils::DeviceRepr;
 use crate::utils::{FileLoc, ParameterDictionary};
+use crate::Arena;
 use anyhow::{anyhow, Result};
-use shared::core::filter::{DeviceFilterSampler, Filter};
+use shared::core::filter::Filter;
 use shared::core::geometry::{Bounds2f, Point2f, Vector2f};
 use shared::filters::*;
-use shared::Float;
+use shared::{Array2D, Float};
 
 pub trait FilterFactory {
-    fn create(name: &str, params: &ParameterDictionary, loc: &FileLoc) -> Result<Filter>;
+    fn create(
+        name: &str,
+        params: &ParameterDictionary,
+        loc: &FileLoc,
+        arena: &Arena,
+    ) -> Result<Filter>;
 }
 
 impl FilterFactory for Filter {
-    fn create(name: &str, params: &ParameterDictionary, loc: &FileLoc) -> Result<Self> {
+    fn create(
+        name: &str,
+        params: &ParameterDictionary,
+        loc: &FileLoc,
+        arena: &Arena,
+    ) -> Result<Self> {
         match name {
             "box" => {
                 let xw = params.get_one_float("xradius", 0.5)?;
@@ -27,7 +35,6 @@ impl FilterFactory for Filter {
                 let yw = params.get_one_float("yradius", 1.5)?;
                 let sigma = params.get_one_float("sigma", 0.5)?;
                 let filter = GaussianFilter::new(Vector2f::new(xw, yw), sigma);
-                Ok(Filter::Gaussian(filter))
             }
             "mitchell" => {
                 let xw = params.get_one_float("xradius", 2.)?;
@@ -54,52 +61,3 @@ impl FilterFactory for Filter {
         }
     }
 }
-
-#[repr(C)]
-#[derive(Clone, Debug, Copy)]
-pub struct FilterSampler {
-    pub domain: Bounds2f,
-    pub distrib: PiecewiseConstant2D,
-    pub f: Array2D<Float>,
-}
-
-impl FilterSampler {
-    pub fn new<F>(radius: Vector2f, func: F) -> Self
-    where
-        F: Fn(Point2f) -> Float,
-    {
-        let domain = Bounds2f::from_points(
-            Point2f::new(-radius.x(), -radius.y()),
-            Point2f::new(radius.x(), radius.y()),
-        );
-
-        let nx = (32.0 * radius.x()) as i32;
-        let ny = (32.0 * radius.y()) as i32;
-        let mut f = Array2D::new_dims(nx, ny);
-        for y in 0..f.y_size() {
-            for x in 0..f.x_size() {
-                let p = domain.lerp(Point2f::new(
-                    (x as Float + 0.5) / f.x_size() as Float,
-                    (y as Float + 0.5) / f.y_size() as Float,
-                ));
-                f[(x as i32, y as i32)] = func(p);
-            }
-        }
-
-        let distrib = PiecewiseConstant2D::new_with_bounds(&f, domain);
-
-        Self { domain, distrib, f }
-    }
-}
-
-impl DeviceRepr for FilterSampler {
-    type Target = DeviceFilterSampler;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> DeviceFilterSampler {
-        DeviceFilterSampler {
-            domain: self.domain,
-            distrib: self.distrib.upload_value(arena),
-            f: self.f.upload_value(arena),
-        }
-    }
-}
diff --git a/src/core/image/metadata.rs b/src/core/image/metadata.rs
index 8fad636..ea926ca 100644
--- a/src/core/image/metadata.rs
+++ b/src/core/image/metadata.rs
@@ -4,8 +4,6 @@ use shared::spectra::RGBColorSpace;
 use shared::utils::math::SquareMatrix;
 use std::collections::HashMap;
 
-// use std::ops::{Deref, DerefMut};
-
 #[derive(Debug, Clone, Default)]
 pub struct ImageChannelDesc {
     pub offset: Vec<usize>,
diff --git a/src/core/image/mod.rs b/src/core/image/mod.rs
index 323f4db..5a9b509 100644
--- a/src/core/image/mod.rs
+++ b/src/core/image/mod.rs
@@ -6,7 +6,7 @@ use shared::Float;
 use shared::Ptr;
 use shared::core::color::{ColorEncoding, ColorEncodingTrait, LINEAR};
 use shared::core::geometry::{Bounds2f, Point2f, Point2i};
-use shared::core::image::{DeviceImage, ImageBase, PixelFormat, Pixels, WrapMode, WrapMode2D};
+use shared::core::image::{Image, ImageBase, PixelFormat, Pixels, WrapMode, WrapMode2D};
 use shared::utils::math::square;
 use smallvec::{SmallVec, smallvec};
 use std::ops::{Deref, DerefMut};
@@ -69,133 +69,66 @@ impl DerefMut for ImageChannelValues {
 }
 
 #[derive(Debug, Clone)]
-pub enum PixelStorage {
-    U8(Vec<u8>),
-    F16(Vec<f16>),
-    F32(Vec<f32>),
-}
-
-impl PixelStorage {
-    pub fn as_pixels(&self) -> Pixels {
-        match self {
-            PixelStorage::U8(vec) => Pixels::new_u8(Ptr::from_raw(vec.as_ptr())),
-            PixelStorage::F16(vec) => Pixels::new_f16(Ptr::from_raw(vec.as_ptr() as *const u16)),
-            PixelStorage::F32(vec) => Pixels::new_f32(Ptr::from_raw(vec.as_ptr())),
-        }
-    }
-
-    pub fn format(&self) -> PixelFormat {
-        match self {
-            PixelStorage::U8(_) => PixelFormat::U8,
-            PixelStorage::F16(_) => PixelFormat::F16,
-            PixelStorage::F32(_) => PixelFormat::F32,
-        }
-    }
-
-    pub fn len(&self) -> usize {
-        match self {
-            PixelStorage::U8(d) => d.len(),
-            PixelStorage::F16(d) => d.len(),
-            PixelStorage::F32(d) => d.len(),
-        }
-    }
-}
-
-#[derive(Debug, Clone)]
-pub struct Image {
-    pub pixels: PixelStorage,
+pub struct HostImage {
+    pub inner: Image,
     pub channel_names: Vec<String>,
-    pub device: DeviceImage,
 }
 
-// impl Deref for Image {
-//     type Target = DeviceImage;
-//     #[inline]
-//     fn deref(&self) -> &Self::Target {
-//         &self.device
-//     }
-// }
-
 #[derive(Debug, Clone)]
 pub struct ImageAndMetadata {
-    pub image: Image,
+    pub image: HostImage,
     pub metadata: ImageMetadata,
 }
 
-impl Image {
-    // Constructors
-    fn from_storage(
-        storage: PixelStorage,
+impl HostImage {
+    pub fn from_u8(
+        data: &[u8],
         resolution: Point2i,
         channel_names: &[impl AsRef<str>],
         encoding: ColorEncoding,
     ) -> Self {
         let n_channels = channel_names.len() as i32;
-        let expected = (resolution.x() * resolution.y()) as usize * n_channels as usize;
-        let channel_names = channel_names
-            .iter()
-            .map(|s| s.as_ref().to_string())
-            .collect();
-        assert_eq!(storage.len(), expected, "Pixel data size mismatch");
-
-        let device = DeviceImage {
-            base: ImageBase {
-                format: storage.format(),
-                encoding,
-                resolution,
-                n_channels,
-            },
-            pixels: storage.as_pixels(),
-        };
-
         Self {
-            pixels: storage,
-            channel_names,
-            device,
+            inner: Image::from_u8(data, resolution, n_channels, encoding),
+            channel_names: channel_names.iter().map(|s| s.as_ref().to_string()).collect(),
         }
     }
 
-    pub fn from_u8(
-        data: Vec<u8>,
+    pub fn from_f32(
+        data: &[f32],
         resolution: Point2i,
         channel_names: &[impl AsRef<str>],
-        encoding: ColorEncoding,
     ) -> Self {
-        Self::from_storage(PixelStorage::U8(data), resolution, channel_names, encoding)
+        let n_channels = channel_names.len() as i32;
+        Self {
+            inner: Image::from_f32(data, resolution, n_channels),
+            channel_names: channel_names.iter().map(|s| s.as_ref().to_string()).collect(),
+        }
     }
 
     pub fn from_f16(
-        data: Vec<half::f16>,
+        data: &[u16],
         resolution: Point2i,
         channel_names: &[impl AsRef<str>],
     ) -> Self {
-        Self::from_storage(PixelStorage::F16(data), resolution, channel_names, LINEAR)
-    }
-
-    pub fn from_f32(
-        data: Vec<f32>,
-        resolution: Point2i,
-        channel_names: &[impl AsRef<str>],
-    ) -> Self {
-        Self::from_storage(PixelStorage::F32(data), resolution, channel_names, LINEAR)
+        let n_channels = channel_names.len() as i32;
+        Self {
+            inner: Image::from_f16(data, resolution, n_channels),
+            channel_names: channel_names.iter().map(|s| s.as_ref().to_string()).collect(),
+        }
     }
 
     pub fn new(
         format: PixelFormat,
         resolution: Point2i,
         channel_names: &[impl AsRef<str>],
-        encoding: Arc<ColorEncoding>,
+        encoding: ColorEncoding,
     ) -> Self {
-        let n_channels = channel_names.len();
-        let pixel_count = (resolution.x() * resolution.y()) as usize * n_channels;
-
-        let storage = match format {
-            PixelFormat::U8 => PixelStorage::U8(vec![0; pixel_count].into()),
-            PixelFormat::F16 => PixelStorage::F16(vec![f16::ZERO; pixel_count].into()),
-            PixelFormat::F32 => PixelStorage::F32(vec![0.0; pixel_count].into()),
-        };
-
-        Self::from_storage(storage, resolution, channel_names, *encoding)
+        let n_channels = channel_names.len() as i32;
+        Self {
+            inner: Image::new(format, resolution, n_channels, encoding),
+            channel_names: channel_names.iter().map(|s| s.as_ref().to_string()).collect(),
+        }
     }
 
     pub fn new_constant(
@@ -204,206 +137,29 @@ impl Image {
         values: &[f32],
     ) -> Self {
         let n_channels = channel_names.len();
-        if values.len() != n_channels {
-            panic!(
-                "Image::new_constant: values length ({}) must match channel count ({})",
-                values.len(),
-                n_channels
-            );
-        }
+        assert_eq!(values.len(), n_channels, "values length must match channel count");
 
         let n_pixels = (resolution.x() * resolution.y()) as usize;
-
         let mut data = Vec::with_capacity(n_pixels * n_channels);
-
         for _ in 0..n_pixels {
             data.extend_from_slice(values);
         }
 
-        Self::from_f32(data, resolution, channel_names)
-    }
-
-    // Access
-    pub fn device(&self) -> &DeviceImage {
-        &self.device
-    }
-
-    pub fn resolution(&self) -> Point2i {
-        self.base.resolution
-    }
-
-    fn n_channels(&self) -> i32 {
-        self.base().n_channels
-    }
-
-    pub fn format(&self) -> PixelFormat {
-        self.base().format
+        Self::from_f32(&data, resolution, channel_names)
     }
 
     pub fn channel_names(&self) -> Vec<&str> {
         self.channel_names.iter().map(|s| s.as_str()).collect()
     }
 
-    pub fn encoding(&self) -> ColorEncoding {
-        self.base().encoding
-    }
-
-    fn pixel_offset(&self, p: Point2i) -> usize {
-        let width = self.resolution().x() as usize;
-        let idx = p.y() as usize * width + p.x() as usize;
-        idx * self.n_channels() as usize
-    }
-
-    // Read
-    pub fn as_f32_slice(&self) -> Option<&[f32]> {
-        match &self.pixels {
-            PixelStorage::F32(data) => Some(data.as_slice()),
-            _ => None,
-        }
-    }
-    pub fn get_channel(&self, p: Point2i, c: i32) -> Float {
-        self.get_channel_with_wrap(p, c, WrapMode::Clamp.into())
-    }
-
-    pub fn get_channel_with_wrap(&self, mut p: Point2i, c: i32, wrap_mode: WrapMode2D) -> Float {
-        if !self.device.base.remap_pixel_coords(&mut p, wrap_mode) {
-            return 0.0;
-        }
-
-        let offset = self.pixel_offset(p) + c as usize;
-
-        match &self.pixels {
-            PixelStorage::U8(data) => self.device.base.encoding.to_linear_scalar(data[offset]),
-            PixelStorage::F16(data) => data[offset].to_f32(),
-            PixelStorage::F32(data) => data[offset],
-        }
-    }
-
-    pub fn get_channels(&self, p: Point2i) -> ImageChannelValues {
-        self.get_channels_with_wrap(p, WrapMode::Clamp.into())
-    }
-
-    pub fn get_channels_with_wrap(
-        &self,
-        mut p: Point2i,
-        wrap_mode: WrapMode2D,
-    ) -> ImageChannelValues {
-        if !self.device.base.remap_pixel_coords(&mut p, wrap_mode) {
-            return ImageChannelValues(smallvec![0.0; self.n_channels() as usize]);
-        }
-
-        let offset = self.pixel_offset(p);
-        let nc = self.n_channels() as usize;
-        let mut values = SmallVec::with_capacity(nc);
-
-        match &self.pixels {
-            PixelStorage::U8(data) => {
-                for i in 0..nc {
-                    values.push(self.device.base.encoding.to_linear_scalar(data[offset + i]));
-                }
-            }
-            PixelStorage::F16(data) => {
-                for i in 0..nc {
-                    values.push(data[offset + i].to_f32());
-                }
-            }
-            PixelStorage::F32(data) => {
-                for i in 0..nc {
-                    values.push(data[offset + i]);
-                }
-            }
-        }
-
-        ImageChannelValues(values)
-    }
-
-    // Write
-    pub fn set_channel(&mut self, p: Point2i, c: i32, mut value: Float) {
-        if value.is_nan() {
-            value = 0.0;
-        }
-
-        let res = self.resolution();
-        if p.x() < 0 || p.x() >= res.x() || p.y() < 0 || p.y() >= res.y() {
-            return;
-        }
-
-        let offset = self.pixel_offset(p) + c as usize;
-
-        match &mut self.pixels {
-            PixelStorage::U8(data) => {
-                data[offset] = self.device.base.encoding.from_linear_scalar(value);
-            }
-            PixelStorage::F16(data) => {
-                data[offset] = f16::from_f32(value);
-            }
-            PixelStorage::F32(data) => {
-                data[offset] = value;
-            }
-        }
-    }
-
-    pub fn set_channels(&mut self, p: Point2i, values: &ImageChannelValues) {
-        for i in 0..values.len() {
-            self.set_channel(p, i.try_into().unwrap(), values[i])
-        }
-    }
-
-    // Descriptions
-    pub fn get_channels_with_desc(
-        &self,
-        p: Point2i,
-        desc: &ImageChannelDesc,
-        wrap_mode: WrapMode2D,
-    ) -> ImageChannelValues {
-        let mut pp = p;
-        if !self.device.base.remap_pixel_coords(&mut pp, wrap_mode) {
-            return ImageChannelValues(smallvec![0.0; desc.offset.len()]);
-        }
-
-        let pixel_offset = self.pixel_offset(pp);
-        let mut values = SmallVec::with_capacity(desc.offset.len());
-
-        match &self.pixels {
-            PixelStorage::U8(data) => {
-                for &c in &desc.offset {
-                    let raw = data[pixel_offset + c];
-                    values.push(self.device.base.encoding.to_linear_scalar(raw));
-                }
-            }
-            PixelStorage::F16(data) => {
-                for &c in &desc.offset {
-                    values.push(data[pixel_offset + c].to_f32());
-                }
-            }
-            PixelStorage::F32(data) => {
-                for &c in &desc.offset {
-                    values.push(data[pixel_offset + c]);
-                }
-            }
-        }
-
-        ImageChannelValues(values)
-    }
-
-    pub fn channel_names_from_desc(&self, desc: &ImageChannelDesc) -> Vec<&str> {
-        desc.offset
-            .iter()
-            .map(|&i| self.channel_names[i].as_str())
-            .collect()
-    }
-
     pub fn get_channel_desc(
         &self,
-        requested_channels: &[impl AsRef<str> + std::fmt::Display],
+        requested: &[impl AsRef<str> + std::fmt::Display],
     ) -> Result<ImageChannelDesc> {
-        let mut offset = Vec::with_capacity(requested_channels.len());
-
-        for req in requested_channels.iter() {
+        let mut offset = Vec::with_capacity(requested.len());
+        for req in requested {
             match self.channel_names.iter().position(|n| n == req.as_ref()) {
-                Some(idx) => {
-                    offset.push(idx);
-                }
+                Some(idx) => offset.push(idx),
                 None => {
                     return Err(anyhow!(
                         "Missing channel '{}'. Available: {:?}",
@@ -413,81 +169,98 @@ impl Image {
                 }
             }
         }
-
         Ok(ImageChannelDesc { offset })
     }
 
     pub fn all_channels_desc(&self) -> ImageChannelDesc {
         ImageChannelDesc {
-            offset: (0..self.n_channels() as usize).collect(),
+            offset: (0..self.inner.n_channels() as usize).collect(),
+        }
+    }
+
+    pub fn channel_names_from_desc(&self, desc: &ImageChannelDesc) -> Vec<&str> {
+        desc.offset.iter().map(|&i| self.channel_names[i].as_str()).collect()
+    }
+
+    pub fn get_channels(&self, p: Point2i) -> ImageChannelValues {
+        self.get_channels_with_wrap(p, WrapMode::Clamp.into())
+    }
+
+    pub fn get_channels_with_wrap(&self, mut p: Point2i, wrap_mode: WrapMode2D) -> ImageChannelValues {
+        if !self.inner.remap_pixel_coords(&mut p, wrap_mode) {
+            return ImageChannelValues(SmallVec::from_elem(0.0, self.inner.n_channels() as usize));
+        }
+        let offset = self.inner.pixel_offset(p);
+        let nc = self.inner.n_channels() as usize;
+        let mut values = SmallVec::with_capacity(nc);
+        for i in 0..nc {
+            values.push(unsafe { self.inner.pixels.read(offset + i, &self.inner.encoding) });
+        }
+        ImageChannelValues(values)
+    }
+
+    pub fn get_channels_with_desc(
+        &self,
+        p: Point2i,
+        desc: &ImageChannelDesc,
+        wrap_mode: WrapMode2D,
+    ) -> ImageChannelValues {
+        let mut pp = p;
+        if !self.inner.remap_pixel_coords(&mut pp, wrap_mode) {
+            return ImageChannelValues(SmallVec::from_elem(0.0, desc.offset.len()));
+        }
+        let pixel_offset = self.inner.pixel_offset(pp);
+        let mut values = SmallVec::with_capacity(desc.offset.len());
+        for &c in &desc.offset {
+            values.push(unsafe { self.inner.pixels.read(pixel_offset + c, &self.inner.encoding) });
+        }
+        ImageChannelValues(values)
+    }
+
+    pub fn set_channels(&mut self, p: Point2i, values: &ImageChannelValues) {
+        for i in 0..values.len() {
+            self.inner.set_channel(p, i as i32, values[i]);
         }
     }
 
     pub fn select_channels(&self, desc: &ImageChannelDesc) -> Self {
-        let new_names: Vec<String> = desc
-            .offset
-            .iter()
-            .map(|&i| self.channel_names[i].clone())
-            .collect();
-
-        let res = self.resolution();
+        let new_names: Vec<String> = desc.offset.iter().map(|&i| self.channel_names[i].clone()).collect();
+        let res = self.inner.resolution();
         let pixel_count = (res.x() * res.y()) as usize;
-        let src_nc = self.n_channels() as usize;
+        let src_nc = self.inner.n_channels() as usize;
         let dst_nc = desc.offset.len();
 
-        let new_storage = match &self.pixels {
-            PixelStorage::U8(src) => {
-                let mut dst = vec![0u8; pixel_count * dst_nc];
-                for i in 0..pixel_count {
-                    for (out_idx, &in_c) in desc.offset.iter().enumerate() {
-                        dst[i * dst_nc + out_idx] = src[i * src_nc + in_c];
-                    }
-                }
-                PixelStorage::U8(dst)
+        // Always produce f32 output for simplicity
+        let mut dst = vec![0.0f32; pixel_count * dst_nc];
+        for i in 0..pixel_count {
+            let src_offset = i * src_nc;
+            for (out_idx, &in_c) in desc.offset.iter().enumerate() {
+                dst[i * dst_nc + out_idx] = unsafe {
+                    self.inner.pixels.read(src_offset + in_c, &self.inner.encoding)
+                };
             }
-            PixelStorage::F16(src) => {
-                let mut dst = vec![f16::ZERO; pixel_count * dst_nc];
-                for i in 0..pixel_count {
-                    for (out_idx, &in_c) in desc.offset.iter().enumerate() {
-                        dst[i * dst_nc + out_idx] = src[i * src_nc + in_c];
-                    }
-                }
-                PixelStorage::F16(dst)
-            }
-            PixelStorage::F32(src) => {
-                let mut dst = vec![0.0f32; pixel_count * dst_nc];
-                for i in 0..pixel_count {
-                    for (out_idx, &in_c) in desc.offset.iter().enumerate() {
-                        dst[i * dst_nc + out_idx] = src[i * src_nc + in_c];
-                    }
-                }
-                PixelStorage::F32(dst)
-            }
-        };
+        }
 
-        Self::from_storage(new_storage, res, &new_names, self.encoding())
+        Self::from_f32(&dst, res, &new_names)
     }
 
     pub fn get_sampling_distribution<F>(&self, dxd_a: F, domain: Bounds2f) -> Array2D<Float>
     where
         F: Fn(Point2f) -> Float + Sync + Send,
     {
-        let width = self.resolution().x();
-        let height = self.resolution().y();
+        let width = self.inner.resolution().x();
+        let height = self.inner.resolution().y();
 
-        let mut dist: Array2D<Float> = Array2D::new_dims(width, height);
+        let mut dist = Array2D::new_dims(width, height);
 
-        dist.values
+        dist.as_mut_slice()
             .par_chunks_mut(width as usize)
             .enumerate()
             .for_each(|(y, row)| {
                 let y = y as i32;
-
                 for (x, out_val) in row.iter_mut().enumerate() {
                     let x = x as i32;
-
                     let value = self.get_channels(Point2i::new(x, y)).average();
-
                     let u = (x as Float + 0.5) / width as Float;
                     let v = (y as Float + 0.5) / height as Float;
                     let p = domain.lerp(Point2f::new(u, v));
@@ -499,29 +272,36 @@ impl Image {
     }
 
     pub fn get_sampling_distribution_uniform(&self) -> Array2D<Float> {
-        let default_domain = Bounds2f::from_points(Point2f::new(0.0, 0.0), Point2f::new(1.0, 1.0));
+        let domain = Bounds2f::from_points(Point2f::new(0.0, 0.0), Point2f::new(1.0, 1.0));
+        self.get_sampling_distribution(|_| 1.0, domain)
+    }
 
-        self.get_sampling_distribution(|_| 1.0, default_domain)
+    pub fn has_any_infinite_pixels(&self) -> bool {
+        self.inner.has_any_infinite_pixels()
+    }
+
+    pub fn has_any_nan_pixels(&self) -> bool {
+        self.inner.has_any_nan_pixels()
     }
 
     pub fn mse(
         &self,
-        desc: ImageChannelDesc,
-        ref_img: &Image,
+        desc: &ImageChannelDesc,
+        ref_img: &HostImage,
         generate_mse_image: bool,
-    ) -> (ImageChannelValues, Option<Image>) {
-        let res = self.resolution();
+    ) -> (ImageChannelValues, Option<HostImage>) {
+        let res = self.inner.resolution();
+        assert_eq!(res, ref_img.inner.resolution());
 
-        let mut sum_se: Vec<f64> = vec![0.; desc.size()];
-        let names_ref = self.channel_names_from_desc(&desc);
         let ref_desc = ref_img
-            .get_channel_desc(&self.channel_names_from_desc(&desc))
-            .expect("Channels not found in image");
-        assert_eq!(res, ref_img.resolution());
+            .get_channel_desc(&self.channel_names_from_desc(desc))
+            .expect("Channels not found in reference image");
 
         let width = res.x() as usize;
         let height = res.y() as usize;
-        let n_channels = desc.offset.len();
+        let n_channels = desc.size();
+
+        let mut sum_se: Vec<f64> = vec![0.0; n_channels];
         let mut mse_pixels = if generate_mse_image {
             vec![0.0f32; width * height * n_channels]
         } else {
@@ -530,18 +310,15 @@ impl Image {
 
         for y in 0..res.y() {
             for x in 0..res.x() {
-                let v =
-                    self.get_channels_with_desc(Point2i::new(x, y), &desc, WrapMode::Clamp.into());
-                let v_ref = self.get_channels_with_desc(
-                    Point2i::new(x, y),
-                    &ref_desc,
-                    WrapMode::Clamp.into(),
+                let v = self.get_channels_with_desc(
+                    Point2i::new(x, y), desc, WrapMode::Clamp.into(),
                 );
-                for c in 0..desc.size() {
+                let v_ref = ref_img.get_channels_with_desc(
+                    Point2i::new(x, y), &ref_desc, WrapMode::Clamp.into(),
+                );
+                for c in 0..n_channels {
                     let se = square(v[c] as f64 - v_ref[c] as f64);
-                    if se.is_infinite() {
-                        continue;
-                    }
+                    if se.is_infinite() { continue; }
                     sum_se[c] += se;
                     if generate_mse_image {
                         let idx = (y as usize * width + x as usize) * n_channels + c;
@@ -556,7 +333,8 @@ impl Image {
             sum_se.iter().map(|&s| (s / pixel_count) as Float).collect();
 
         let mse_image = if generate_mse_image {
-            Some(Image::new(PixelFormat::F32, res, &names_ref, LINEAR.into()))
+            let names = self.channel_names_from_desc(desc);
+            Some(HostImage::from_f32(&mse_pixels, res, &names))
         } else {
             None
         };
@@ -564,53 +342,11 @@ impl Image {
         (ImageChannelValues(mse_values), mse_image)
     }
 
-    pub fn update_view_pointers(&mut self) {
-        self.device.pixels = match &self.pixels {
-            PixelStorage::U8(vec) => Pixels::new_u8(Ptr::from_raw(vec.as_ptr())),
-            PixelStorage::F16(vec) => Pixels::new_f16(Ptr::from_raw(vec.as_ptr() as *const u16)),
-            PixelStorage::F32(vec) => Pixels::new_f32(Ptr::from_raw(vec.as_ptr())),
-        };
+    pub fn image(&self) -> &Image {
+        &self.inner
     }
 
-    pub fn has_any_infinite_pixels(&self) -> bool {
-        if self.format() == PixelFormat::F32 {
-            return false;
-        }
-
-        for y in 0..self.resolution().y() {
-            for x in 0..self.resolution().x() {
-                for c in 0..self.n_channels() {
-                    if self.get_channel(Point2i::new(x, y), c).is_infinite() {
-                        return true;
-                    }
-                }
-            }
-        }
-        return false;
-    }
-
-    pub fn has_any_nan_pixels(&self) -> bool {
-        if self.format() == PixelFormat::F32 {
-            return false;
-        }
-
-        for y in 0..self.resolution().y() {
-            for x in 0..self.resolution().x() {
-                for c in 0..self.n_channels() {
-                    if self.get_channel(Point2i::new(x, y), c).is_nan() {
-                        return true;
-                    }
-                }
-            }
-        }
-        return false;
-    }
-}
-
-impl std::ops::Deref for Image {
-    type Target = DeviceImage;
-
-    fn deref(&self) -> &DeviceImage {
-        &self.device
+    pub fn into_image(self) -> Image {
+        self.inner
     }
 }
diff --git a/src/core/image/ops.rs b/src/core/image/ops.rs
index 96fade6..c50149b 100644
--- a/src/core/image/ops.rs
+++ b/src/core/image/ops.rs
@@ -1,12 +1,12 @@
-use super::Image;
-use crate::core::image::PixelStorage;
+use super::HostImage;
 use crate::core::image::pixel::PixelStorageTrait;
+use crate::core::image::PixelStorage;
 use rayon::prelude::*;
-use shared::Float;
 use shared::core::color::ColorEncoding;
 use shared::core::geometry::{Bounds2i, Point2i};
 use shared::core::image::{PixelFormat, WrapMode, WrapMode2D};
 use shared::utils::math::windowed_sinc;
+use shared::Float;
 use std::sync::{Arc, Mutex};
 
 #[derive(Debug, Clone, Copy)]
@@ -15,71 +15,105 @@ pub struct ResampleWeight {
     pub weight: [Float; 4],
 }
 
-impl Image {
+impl HostImage {
     pub fn flip_y(&mut self) {
-        let res = self.resolution();
-        let nc = self.n_channels() as usize;
+        let res = self.inner.resolution;
+        let nc = self.inner.n_channels as usize;
 
-        match &mut self.pixels {
-            PixelStorage::U8(d) => flip_y_kernel(d, res, nc),
-            PixelStorage::F16(d) => flip_y_kernel(d, res, nc),
-            PixelStorage::F32(d) => flip_y_kernel(d, res, nc),
+        match self.inner.format {
+            PixelFormat::U8 => flip_y_kernel(self.inner.pixels.as_u8_mut(), res, nc),
+            PixelFormat::F16 => flip_y_kernel(self.inner.pixels.as_f16_mut(), res, nc),
+            PixelFormat::F32 => flip_y_kernel(self.inner.pixels.as_f32_slice_mut(), res, nc),
         }
     }
 
-    pub fn crop(&self, bounds: Bounds2i) -> Image {
-        let res = self.resolution();
-        let n_channels = self.n_channels() as usize;
-
+    pub fn crop(&self, bounds: Bounds2i) -> HostImage {
+        let n_channels = self.inner.n_channels as usize;
         let new_res = Point2i::new(
             bounds.p_max.x() - bounds.p_min.x(),
             bounds.p_max.y() - bounds.p_min.y(),
         );
 
-        let mut new_image = Image::new(
-            self.format(),
-            new_res,
-            &self.channel_names,
-            self.encoding().into(),
-        );
+        let new_names = self.channel_names.clone();
+        let mut new_image =
+            HostImage::new(self.inner.format, new_res, &new_names, self.inner.encoding);
 
-        match (&self.pixels, &mut new_image.pixels) {
-            (PixelStorage::U8(src), PixelStorage::U8(dst)) => {
-                crop_kernel(src, dst, res, bounds, n_channels)
-            }
-            (PixelStorage::F16(src), PixelStorage::F16(dst)) => {
-                crop_kernel(src, dst, res, bounds, n_channels)
-            }
-            (PixelStorage::F32(src), PixelStorage::F32(dst)) => {
-                crop_kernel(src, dst, res, bounds, n_channels)
-            }
-            _ => panic!("Format mismatch in crop"),
+        match self.inner.format {
+            PixelFormat::U8 => crop_kernel(
+                self.inner.pixels.as_u8(),
+                new_image.inner.pixels.as_u8_mut(),
+                self.inner.resolution,
+                bounds,
+                n_channels,
+            ),
+            PixelFormat::F16 => crop_kernel(
+                self.inner.pixels.as_f16(),
+                new_image.inner.pixels.as_f16_mut(),
+                self.inner.resolution,
+                bounds,
+                n_channels,
+            ),
+            PixelFormat::F32 => crop_kernel(
+                self.inner.pixels.as_f32_slice(),
+                new_image.inner.pixels.as_f32_slice_mut(),
+                self.inner.resolution,
+                bounds,
+                n_channels,
+            ),
         }
 
         new_image
     }
 
     pub fn copy_rect_out(&self, extent: Bounds2i, buf: &mut [Float], wrap: WrapMode2D) {
-        match &self.pixels {
-            PixelStorage::U8(d) => copy_rect_out_kernel(d, self, extent, buf, wrap),
-            PixelStorage::F16(d) => copy_rect_out_kernel(d, self, extent, buf, wrap),
-            PixelStorage::F32(d) => copy_rect_out_kernel(d, self, extent, buf, wrap),
+        match self.inner.format {
+            PixelFormat::U8 => {
+                copy_rect_out_kernel(self.inner.pixels.as_u8(), self, extent, buf, wrap)
+            }
+            PixelFormat::F16 => {
+                copy_rect_out_kernel(self.inner.pixels.as_f16(), self, extent, buf, wrap)
+            }
+            PixelFormat::F32 => {
+                copy_rect_out_kernel(self.inner.pixels.as_f32_slice(), self, extent, buf, wrap)
+            }
         }
     }
 
     pub fn copy_rect_in(&mut self, extent: Bounds2i, buf: &[Float]) {
-        let res = self.resolution();
-        let n_channels = self.n_channels() as usize;
-        let encoding = self.encoding();
+        let res = self.inner.resolution;
+        let n_channels = self.inner.n_channels as usize;
+        let encoding = self.inner.encoding;
+        let format = self.inner.format;
 
-        match &mut self.pixels {
-            PixelStorage::U8(d) => copy_rect_in_kernel(d, res, n_channels, encoding, extent, buf),
-            PixelStorage::F16(d) => copy_rect_in_kernel(d, res, n_channels, encoding, extent, buf),
-            PixelStorage::F32(d) => copy_rect_in_kernel(d, res, n_channels, encoding, extent, buf),
+        match format {
+            PixelFormat::U8 => copy_rect_in_kernel(
+                self.inner.pixels.as_u8_mut(),
+                res,
+                n_channels,
+                encoding,
+                extent,
+                buf,
+            ),
+            PixelFormat::F16 => copy_rect_in_kernel(
+                self.inner.pixels.as_f16_mut(),
+                res,
+                n_channels,
+                encoding,
+                extent,
+                buf,
+            ),
+            PixelFormat::F32 => copy_rect_in_kernel(
+                self.inner.pixels.as_f32_slice_mut(),
+                res,
+                n_channels,
+                encoding,
+                extent,
+                buf,
+            ),
         }
     }
 
-    pub fn float_resize_up(&self, new_res: Point2i, wrap_mode: WrapMode2D) -> Image {
+    pub fn float_resize_up(&self, new_res: Point2i, wrap_mode: WrapMode2D) -> HostImage {
         let res = self.resolution();
         assert!(new_res.x() >= res.x() && new_res.y() >= res.y());
         assert!(
@@ -87,7 +121,7 @@ impl Image {
             "ResizeUp requires Float format"
         );
 
-        let resampled_image = Arc::new(Mutex::new(Image::new(
+        let resampled_image = Arc::new(Mutex::new(HostImage::new(
             PixelFormat::F32,
             new_res,
             &self.channel_names,
@@ -132,7 +166,7 @@ impl Image {
             .unwrap()
     }
 
-    pub fn generate_pyramid(base: Image, _wrap: WrapMode) -> Vec<Image> {
+    pub fn generate_pyramid(base: HostImage, _wrap: WrapMode) -> Vec<HostImage> {
         let mut levels = vec![base];
         let internal_wrap = WrapMode2D {
             uv: [WrapMode::Clamp; 2],
@@ -146,17 +180,32 @@ impl Image {
             }
 
             let new_res = Point2i::new((old.x() / 2).max(1), (old.y() / 2).max(1));
-            let mut next = Image::new(
-                prev.format(),
+            let mut next = HostImage::new(
+                prev.inner.format,
                 new_res,
                 &prev.channel_names,
-                prev.encoding().into(),
+                prev.inner.encoding,
             );
 
-            match &mut next.pixels {
-                PixelStorage::U8(d) => downsample_kernel(d, new_res, prev, internal_wrap),
-                PixelStorage::F16(d) => downsample_kernel(d, new_res, prev, internal_wrap),
-                PixelStorage::F32(d) => downsample_kernel(d, new_res, prev, internal_wrap),
+            match next.inner.format {
+                PixelFormat::U8 => downsample_kernel(
+                    next.inner.pixels.as_u8_mut(),
+                    new_res,
+                    &prev.inner,
+                    internal_wrap,
+                ),
+                PixelFormat::F16 => downsample_kernel(
+                    next.inner.pixels.as_f16_mut(),
+                    new_res,
+                    &prev.inner,
+                    internal_wrap,
+                ),
+                PixelFormat::F32 => downsample_kernel(
+                    next.inner.pixels.as_f32_slice_mut(),
+                    new_res,
+                    &prev.inner,
+                    internal_wrap,
+                ),
             }
             levels.push(next);
         }
@@ -202,7 +251,7 @@ fn crop_kernel<T: PixelStorageTrait>(
 
 fn copy_rect_out_kernel<T: PixelStorageTrait>(
     src: &[T],
-    image: &Image,
+    image: &HostImage,
     extent: Bounds2i,
     buf: &mut [Float],
     wrap: WrapMode2D,
diff --git a/src/core/scene/scene.rs b/src/core/scene/scene.rs
index e6bb424..e7bd4c0 100644
--- a/src/core/scene/scene.rs
+++ b/src/core/scene/scene.rs
@@ -11,8 +11,8 @@ use crate::core::shape::{ShapeFactory, ShapeWithContext};
 use crate::core::texture::{FloatTexture, SpectrumTexture};
 use crate::utils::parallel::{run_async, AsyncJob};
 use crate::utils::parameters::{NamedTextures, ParameterDictionary, TextureParameterDictionary};
-use crate::utils::{resolve_filename, Upload};
-use crate::{Arena, FileLoc};
+use crate::utils::resolve_filename;
+use crate::{Arena, DeviceRepr, FileLoc};
 use anyhow::{anyhow, Result};
 use parking_lot::Mutex;
 use shared::core::camera::{Camera, CameraTransform};
@@ -121,7 +121,7 @@ impl BasicScene {
         sampler: SceneEntity,
         integ: SceneEntity,
         accel: SceneEntity,
-        arena: Arc<Arena>,
+        arena: &Arena,
     ) -> Result<()> {
         *self.integrator.lock() = Some(integ);
         *self.accelerator.lock() = Some(accel);
@@ -130,7 +130,7 @@ impl BasicScene {
             *self.film_colorspace.lock() = Some(Arc::clone(cs));
         }
 
-        let filter = Filter::create(&filter.name, &filter.parameters, &filter.loc)
+        let filter = Filter::create(&filter.name, &filter.parameters, &filter.loc, &arena)
             .map_err(|e| anyhow!("Failed to create filter: {}", e))?;
         let shutter_close = camera.base.parameters.get_one_float("shutterclose", 1.)?;
         let shutter_open = camera.base.parameters.get_one_float("shutteropen", 0.)?;
diff --git a/src/core/texture.rs b/src/core/texture.rs
index 40e709a..56d7eca 100644
--- a/src/core/texture.rs
+++ b/src/core/texture.rs
@@ -1,21 +1,21 @@
 use crate::textures::*;
-use crate::utils::mipmap::MIPMap;
-use crate::utils::mipmap::MIPMapFilterOptions;
-use crate::utils::{Arena, FileLoc, TextureParameterDictionary};
-use anyhow::{Result, anyhow};
+use crate::utils::mipmap::{MIPMapFilterOptions, MIPMap};
+use crate::utils::TextureParameterDictionary;
+use crate::{Arena, Device, FileLoc, DeviceRepr};
+use anyhow::{anyhow, Result};
 use enum_dispatch::enum_dispatch;
-use shared::Float;
 use shared::core::color::ColorEncoding;
 use shared::core::geometry::Vector3f;
 use shared::core::image::WrapMode;
 use shared::core::texture::SpectrumType;
 use shared::core::texture::{
     CylindricalMapping, PlanarMapping, SphericalMapping, TextureEvalContext, TextureMapping2D,
-    UVMapping,
+    UVMapping, GPUFloatTexture, GPUSpectrumTexture
 };
 use shared::spectra::{SampledSpectrum, SampledWavelengths};
 use shared::textures::*;
 use shared::utils::Transform;
+use shared::Float;
 use std::collections::HashMap;
 use std::sync::{Arc, Mutex, OnceLock};
 
@@ -29,20 +29,49 @@ pub trait SpectrumTextureTrait {
     fn evaluate(&self, _ctx: &TextureEvalContext, _lambda: &SampledWavelengths) -> SampledSpectrum;
 }
 
-#[enum_dispatch(FloatTextureTrait)]
-#[derive(Debug, Clone)]
+#[derive(Clone, Debug, Device)]
+#[device(name = "GPUFloatTexture")]
 pub enum FloatTexture {
+    #[device(clone)]
     Constant(FloatConstantTexture),
-    Image(FloatImageTexture),
-    Mix(FloatMixTexture),
-    DirectionMix(FloatDirectionMixTexture),
-    Scaled(FloatScaledTexture),
-    Bilerp(FloatBilerpTexture),
+
+    #[device(clone)]
     Checkerboard(FloatCheckerboardTexture),
+
+    #[device(clone)]
     Dots(FloatDotsTexture),
+
+    #[device(clone)]
     FBm(FBmTexture),
+
+    #[device(clone)]
     Windy(WindyTexture),
+
+    #[device(clone)]
     Wrinkled(WrinkledTexture),
+
+    Scaled(FloatScaledTexture),
+
+    Mix(FloatMixTexture),
+
+    DirectionMix(FloatDirectionMixTexture),
+
+    #[device(custom = "upload_image", variant_type = "GPUFloatImageTexture")]
+    Image(FloatImageTexture),
+
+    #[device(clone)]
+    Bilerp(FloatBilerpTexture),
+}
+
+impl FloatTexture {
+    fn upload_image(inner: &FloatImageTexture, _arena: &Arena) -> GPUFloatImageTexture {
+        GPUFloatImageTexture {
+            mapping: inner.base.mapping,
+            tex_obj: inner.base.mipmap.texture_object(),
+            scale: inner.base.scale,
+            invert: inner.base.invert,
+        }
+    }
 }
 
 impl Default for FloatTexture {
@@ -95,21 +124,58 @@ impl FloatTexture {
     }
 }
 
-#[derive(Clone, Debug)]
-#[enum_dispatch(SpectrumTextureTrait)]
+#[derive(Clone, Debug, Device)]
+#[device(name = "GPUSpectrumTexture")]
 pub enum SpectrumTexture {
-    // RGBConstant(RGBConstantTexture),
-    // RGBReflectanceConstant(RGBReflectanceConstantTexture),
+    #[device(clone)]
     Constant(SpectrumConstantTexture),
-    Bilerp(SpectrumBilerpTexture),
+
+    #[device(clone)]
     Checkerboard(SpectrumCheckerboardTexture),
-    Image(SpectrumImageTexture),
-    Marble(MarbleTexture),
-    Mix(SpectrumMixTexture),
-    DirectionMix(SpectrumDirectionMixTexture),
+
+    #[device(clone)]
     Dots(SpectrumDotsTexture),
-    // Ptex(SpectrumPtexTexture),
+
+    #[device(
+        custom = "upload_spectrum_image",
+        variant_type = "GPUSpectrumImageTexture"
+    )]
+    Image(SpectrumImageTexture),
+
+    #[device(clone)]
+    Bilerp(SpectrumBilerpTexture),
+
     Scaled(SpectrumScaledTexture),
+
+    #[device(clone)]
+    Marble(MarbleTexture),
+
+    Mix(SpectrumMixTexture),
+
+    DirectionMix(SpectrumDirectionMixTexture),
+}
+
+impl SpectrumTexture {
+    fn upload_spectrum_image(
+        inner: &SpectrumImageTexture,
+        arena: &Arena,
+    ) -> GPUSpectrumImageTexture {
+        let tex_obj = arena.get_texture_object(&inner.base.mipmap);
+        GPUSpectrumImageTexture {
+            mapping: inner.base.mapping,
+            tex_obj,
+            scale: inner.base.scale,
+            invert: inner.base.invert,
+            is_single_channel: inner.base.mipmap.is_single_channel(),
+            color_space: inner
+                .base
+                .mipmap
+                .color_space
+                .clone()
+                .unwrap_or_else(crate::spectra::default_colorspace),
+            spectrum_type: inner.spectrum_type,
+        }
+    }
 }
 
 pub trait CreateSpectrumTexture {
diff --git a/src/films/gbuffer.rs b/src/films/gbuffer.rs
index dd74009..a3584d1 100644
--- a/src/films/gbuffer.rs
+++ b/src/films/gbuffer.rs
@@ -35,7 +35,7 @@ impl GBufferFilmHost {
             base: base.clone(),
             output_from_render: *output_from_render,
             apply_inverse,
-            pixels: pixels.device,
+            pixels,
             colorspace: colorspace.clone(),
             max_component_value,
             write_fp16,
diff --git a/src/films/spectral.rs b/src/films/spectral.rs
index ea30234..80b8372 100644
--- a/src/films/spectral.rs
+++ b/src/films/spectral.rs
@@ -9,13 +9,13 @@ use shared::core::film::{FilmBase, SpectralFilm, SpectralPixel};
 use shared::core::filter::FilterTrait;
 use shared::spectra::{LAMBDA_MAX, LAMBDA_MIN, RGBColorSpace};
 use shared::utils::AtomicFloat;
-use shared::utils::containers::DeviceArray2D;
+use shared::utils::containers::Array2D;
 use shared::utils::math::SquareMatrix;
 use std::path::Path;
 use std::sync::Arc;
 
 struct SpectralFilmStorage {
-    pixels: DeviceArray2D<SpectralPixel>,
+    pixels: Array2D<SpectralPixel>,
     bucket_sums: Vec<f64>,
     weight_sums: Vec<f64>,
     bucket_splats: Vec<AtomicFloat>,
@@ -72,7 +72,7 @@ impl SpectralFilmHost {
             filter_integral: base.filter.integral(),
             output_rgbf_from_sensor_rgb: SquareMatrix::identity(),
 
-            pixels: DeviceArray2D {
+            pixels: Array2D {
                 values: pixels.values.as_mut_ptr(),
                 extent: base.pixel_bounds,
                 stride: base.pixel_bounds.p_max.x() - base.pixel_bounds.p_min.x(),
diff --git a/src/filters/boxf.rs b/src/filters/boxf.rs
deleted file mode 100644
index 2af62d1..0000000
--- a/src/filters/boxf.rs
+++ /dev/null
@@ -1 +0,0 @@
-pub trait BoxFilterCreator {}
diff --git a/src/filters/gaussian.rs b/src/filters/gaussian.rs
deleted file mode 100644
index 0267a54..0000000
--- a/src/filters/gaussian.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-use crate::core::filter::CreateFilterSampler;
-use shared::Float;
-use shared::core::filter::FilterSampler;
-use shared::core::geometry::{Point2f, Vector2f};
-use shared::filters::GaussianFilter;
-use shared::utils::math::gaussian;
-
-#[derive(Clone, Debug, Device)]
-#[device(name = "GaussianFilter")]
-pub struct GaussianFilterHost {
-    pub radius: Vector2f,
-    pub sigma: Float,
-    pub exp_x: Float,
-    pub exp_y: Float,
-    #[device(flatten)]
-    pub sampler: FilterSampler,
-}
-
-impl GaussianFilterHost {
-    pub fn new(radius: Vector2f, sigma: Float) -> Self {
-        let exp_x = gaussian(radius.x(), 0.0, sigma);
-        let exp_y = gaussian(radius.y(), 0.0, sigma);
-        let sampler = FilterSampler::new(radius, move |p: Point2f| {
-            let gx = (gaussian(p.x(), 0.0, sigma) - exp_x).max(0.0);
-            let gy = (gaussian(p.y(), 0.0, sigma) - exp_y).max(0.0);
-            gx * gy
-        });
-        Self { radius, sigma, exp_x, exp_y, sampler }
-    }
-}
diff --git a/src/filters/lanczos.rs b/src/filters/lanczos.rs
deleted file mode 100644
index 4497bf0..0000000
--- a/src/filters/lanczos.rs
+++ /dev/null
@@ -1,40 +0,0 @@
-use crate::core::filter::CreateFilterSampler;
-use rand::Rng;
-use shared::Float;
-use shared::core::filter::FilterSampler;
-use shared::core::geometry::{Point2f, Vector2f};
-use shared::filters::LanczosSincFilter;
-use shared::utils::math::{lerp, windowed_sinc};
-
-#[derive(Clone, Debug, Device)]
-#[device(name = "LanczosSincFilter")]
-pub struct LanczosSincFilterHost {
-    pub radius: Vector2f,
-    pub tau: Float,
-    #[device(flatten)]
-    pub sampler: FilterSampler,
-}
-
-impl LanczosSincFilterHost {
-    pub fn new(radius: Vector2f, tau: Float) -> Self {
-        let sampler = FilterSampler::new(radius, move |p: Point2f| {
-            windowed_sinc(p.x(), radius.x(), tau) * windowed_sinc(p.y(), radius.y(), tau)
-        });
-        Self { radius, tau, sampler }
-    }
-}
-
-fn windowed_sinc(x: Float, radius: Float, tau: Float) -> Float {
-    use std::f32::consts::PI;
-    let x = x.abs();
-    if x > radius {
-        return 0.0;
-    }
-    if x < 1e-5 {
-        1.0
-    } else {
-        let xpi = x * PI;
-        let xpit = xpi * tau;
-        (xpi.sin() / xpi) * (xpit.sin() / xpit)
-    }
-}
diff --git a/src/filters/mitchell.rs b/src/filters/mitchell.rs
deleted file mode 100644
index a8b2626..0000000
--- a/src/filters/mitchell.rs
+++ /dev/null
@@ -1,42 +0,0 @@
-use crate::core::filter::CreateFilterSampler;
-use shared::Float;
-use shared::core::filter::FilterSampler;
-use shared::core::geometry::{Point2f, Vector2f};
-use shared::filters::MitchellFilter;
-
-#[derive(Clone, Debug, Device)]
-#[device(name = "MitchellFilter")]
-pub struct MitchellFilterHost {
-    pub radius: Vector2f,
-    pub b: Float,
-    pub c: Float,
-    #[device(flatten)]
-    pub sampler: FilterSampler,
-}
-
-impl MitchellFilterHost {
-    pub fn new(radius: Vector2f, b: Float, c: Float) -> Self {
-        let sampler = FilterSampler::new(radius, move |p: Point2f| {
-            mitchell_1d(p.x() / radius.x(), b, c) * mitchell_1d(p.y() / radius.y(), b, c)
-        });
-        Self { radius, b, c, sampler }
-    }
-}
-
-fn mitchell_1d(x: Float, b: Float, c: Float) -> Float {
-    let x = (2.0 * x).abs();
-    if x <= 1.0 {
-        ((12.0 - 9.0 * b - 6.0 * c) * x * x * x
-            + (-18.0 + 12.0 * b + 6.0 * c) * x * x
-            + (6.0 - 2.0 * b))
-            / 6.0
-    } else if x <= 2.0 {
-        ((-b - 6.0 * c) * x * x * x
-            + (6.0 * b + 30.0 * c) * x * x
-            + (-12.0 * b - 48.0 * c) * x
-            + (8.0 * b + 24.0 * c))
-            / 6.0
-    } else {
-        0.0
-    }
-}
diff --git a/src/filters/mod.rs b/src/filters/mod.rs
deleted file mode 100644
index e4abe99..0000000
--- a/src/filters/mod.rs
+++ /dev/null
@@ -1,11 +0,0 @@
-pub mod boxf;
-pub mod gaussian;
-pub mod lanczos;
-pub mod mitchell;
-pub mod triangle;
-
-pub use boxf::*;
-pub use gaussian::*;
-pub use lanczos::*;
-pub use mitchell::*;
-// pub use triangle::*;
diff --git a/src/filters/triangle.rs b/src/filters/triangle.rs
deleted file mode 100644
index 8b13789..0000000
--- a/src/filters/triangle.rs
+++ /dev/null
@@ -1 +0,0 @@
-
diff --git a/src/lib.rs b/src/lib.rs
index 8c0cf49..c84be56 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -17,4 +17,4 @@ pub mod utils;
 #[cfg(feature = "cuda")]
 pub mod gpu;
 
-pub use utils::{Arena, FileLoc, ParameterDictionary};
+pub use utils::{Arena, Device, DeviceRepr, FileLoc, ParameterDictionary};
diff --git a/src/lights/diffuse.rs b/src/lights/diffuse.rs
index 220b38b..1edd25b 100644
--- a/src/lights/diffuse.rs
+++ b/src/lights/diffuse.rs
@@ -4,8 +4,9 @@ use crate::core::image::{Image, ImageIO};
 use crate::core::light::lookup_spectrum;
 use crate::core::spectrum::spectrum_to_photometric;
 use crate::core::texture::FloatTexture;
-use crate::utils::{Arena, FileLoc, ParameterDictionary, Upload, resolve_filename};
-use anyhow::{Result, anyhow};
+use crate::utils::resolve_filename;
+use crate::{Arena, DeviceRepr, FileLoc, ParameterDictionary};
+use anyhow::{anyhow, Result};
 use shared::core::geometry::Point2i;
 use shared::core::light::{Light, LightBase, LightType};
 use shared::core::medium::{Medium, MediumInterface};
@@ -110,7 +111,7 @@ pub fn create(
         _ => false,
     };
 
-    let (light_type, stored_alpha) = if is_constant_zero {
+    let (light_type, _) = if is_constant_zero {
         (LightType::DeltaPosition, None)
     } else {
         (LightType::Area, Some(alpha))
diff --git a/src/lights/goniometric.rs b/src/lights/goniometric.rs
index a0dd2a8..cec54f3 100644
--- a/src/lights/goniometric.rs
+++ b/src/lights/goniometric.rs
@@ -5,7 +5,8 @@ use crate::core::light::lookup_spectrum;
 use crate::core::spectrum::spectrum_to_photometric;
 use crate::core::texture::FloatTexture;
 use crate::utils::sampling::PiecewiseConstant2D;
-use crate::utils::{Arena, FileLoc, ParameterDictionary, Upload, resolve_filename};
+use crate::utils::resolve_filename;
+use crate::{Arena, FileLoc, ParameterDictionary, DeviceRepr};
 use anyhow::{Result, anyhow};
 use shared::core::geometry::Point2i;
 use shared::core::light::{Light, LightBase, LightType};
@@ -80,7 +81,7 @@ pub fn create(
     ];
     let t =
         Transform::from_flat(&swap_yz).expect("Could not create transform for GoniometricLight");
-    let final_render_from_light = render_from_light * t;
+    let _final_render_from_light = render_from_light * t;
 
     let mi = match medium {
         Some(m) => {
diff --git a/src/lights/infinite.rs b/src/lights/infinite.rs
index 2b6d057..020e6c8 100644
--- a/src/lights/infinite.rs
+++ b/src/lights/infinite.rs
@@ -1,15 +1,15 @@
-use crate::core::image::{Image, ImageIO};
+use crate::core::image::{HostImage, ImageIO};
 use crate::core::light::lookup_spectrum;
 use crate::core::spectrum::spectrum_to_photometric;
 use crate::spectra::get_spectra_context;
+use crate::utils::resolve_filename;
 use crate::utils::sampling::{PiecewiseConstant2D, WindowedPiecewiseConstant2D};
-use crate::utils::{resolve_filename, FileLoc, ParameterDictionary, Upload};
-use crate::Arena;
+use crate::{Arena, DeviceRepr, FileLoc, ParameterDictionary};
 use anyhow::{anyhow, Result};
 use rayon::prelude::*;
 use shared::core::camera::CameraTransform;
 use shared::core::geometry::{cos_theta, Bounds2f, Frame, Point2f, Point2i, Point3f, VectorLike};
-use shared::core::image::{DeviceImage, WrapMode};
+use shared::core::image::WrapMode;
 use shared::core::light::{Light, LightBase, LightType};
 use shared::core::medium::{Medium, MediumInterface};
 use shared::core::spectrum::Spectrum;
@@ -17,19 +17,18 @@ use shared::core::texture::SpectrumType;
 use shared::lights::{ImageInfiniteLight, PortalInfiniteLight, UniformInfiniteLight};
 use shared::spectra::{DenselySampledSpectrum, RGBColorSpace};
 use shared::utils::math::{equal_area_sphere_to_square, equal_area_square_to_sphere};
-use shared::utils::sampling::{DevicePiecewiseConstant2D, DeviceWindowedPiecewiseConstant2D};
-use shared::utils::{Ptr, Transform};
-use shared::{Float, PI};
+use shared::utils::sampling::{PiecewiseConstant2D, WindowedPiecewiseConstant2D};
+use shared::{Float, Ptr, Transform, PI};
 use std::path::Path;
 
 pub trait CreateImageInfiniteLight {
     fn new(
         render_from_light: Transform,
         scale: Float,
-        image: Ptr<DeviceImage>,
+        image: Ptr<HostImage>,
         image_color_space: Ptr<RGBColorSpace>,
-        distrib: Ptr<DevicePiecewiseConstant2D>,
-        compensated_distrib: Ptr<DevicePiecewiseConstant2D>,
+        distrib: Ptr<PiecewiseConstant2D>,
+        compensated_distrib: Ptr<PiecewiseConstant2D>,
     ) -> Self;
 }
 
@@ -37,10 +36,10 @@ impl CreateImageInfiniteLight for ImageInfiniteLight {
     fn new(
         render_from_light: Transform,
         scale: Float,
-        image: Ptr<DeviceImage>,
+        image: Ptr<HostImage>,
         image_color_space: Ptr<RGBColorSpace>,
-        distrib: Ptr<DevicePiecewiseConstant2D>,
-        compensated_distrib: Ptr<DevicePiecewiseConstant2D>,
+        distrib: Ptr<PiecewiseConstant2D>,
+        compensated_distrib: Ptr<PiecewiseConstant2D>,
     ) -> Self {
         let base = LightBase::new(
             LightType::Infinite,
@@ -64,11 +63,11 @@ pub trait CreatePortalInfiniteLight {
     fn new(
         render_from_light: Transform,
         scale: Float,
-        image: Ptr<DeviceImage>,
+        image: Ptr<HostImage>,
         image_color_space: Ptr<RGBColorSpace>,
         portal: [Point3f; 4],
         portal_frame: Frame,
-        distribution: Ptr<DeviceWindowedPiecewiseConstant2D>,
+        distribution: Ptr<WindowedPiecewiseConstant2D>,
     ) -> Self;
 }
 
@@ -76,11 +75,11 @@ impl CreatePortalInfiniteLight for PortalInfiniteLight {
     fn new(
         render_from_light: Transform,
         scale: Float,
-        image: Ptr<DeviceImage>,
+        image: Ptr<HostImage>,
         image_color_space: Ptr<RGBColorSpace>,
         portal: [Point3f; 4],
         portal_frame: Frame,
-        distribution: Ptr<DeviceWindowedPiecewiseConstant2D>,
+        distribution: Ptr<WindowedPiecewiseConstant2D>,
     ) -> Self {
         let base = LightBase::new(
             LightType::Infinite,
diff --git a/src/lights/projection.rs b/src/lights/projection.rs
index 522983f..650d781 100644
--- a/src/lights/projection.rs
+++ b/src/lights/projection.rs
@@ -2,7 +2,8 @@ use crate::core::image::{Image, ImageIO};
 use crate::core::spectrum::spectrum_to_photometric;
 use crate::core::texture::FloatTexture;
 use crate::utils::sampling::PiecewiseConstant2D;
-use crate::utils::{Arena, FileLoc, ParameterDictionary, Upload, resolve_filename};
+use crate::utils::resolve_filename;
+use crate::{Arena, DeviceRepr, FileLoc, ParameterDictionary};
 use anyhow::{Result, anyhow};
 use shared::Float;
 use shared::core::geometry::{
@@ -77,7 +78,7 @@ pub fn create(
     }
 
     let flip = Transform::scale(1., -1., 1.);
-    let render_from_light_flip = render_from_light * flip;
+    let _render_from_light_flip = render_from_light * flip;
 
     let mi = match medium {
         Some(m) => {
diff --git a/src/materials/coated.rs b/src/materials/coated.rs
index 5db529f..c21511c 100644
--- a/src/materials/coated.rs
+++ b/src/materials/coated.rs
@@ -3,8 +3,9 @@ use crate::core::material::CreateMaterial;
 use crate::core::texture::SpectrumTexture;
 use crate::globals::get_options;
 use crate::spectra::data::get_named_spectrum;
-use crate::utils::{Arena, FileLoc, TextureParameterDictionary, Upload};
-use anyhow::{Result, bail};
+use crate::utils::TextureParameterDictionary;
+use crate::{Arena, DeviceRepr, FileLoc};
+use anyhow::{bail, Result};
 use shared::core::material::Material;
 use shared::core::spectrum::Spectrum;
 use shared::core::texture::SpectrumType;
diff --git a/src/materials/complex.rs b/src/materials/complex.rs
index 925b439..1ed7c93 100644
--- a/src/materials/complex.rs
+++ b/src/materials/complex.rs
@@ -2,15 +2,14 @@ use crate::core::image::Image;
 use crate::core::material::CreateMaterial;
 use crate::core::texture::SpectrumTexture;
 use crate::spectra::get_colorspace_device;
-use crate::utils::{Arena, FileLoc, TextureParameterDictionary, Upload};
+use crate::{Arena, DeviceRepr, FileLoc};
+use crate::utils::TextureParameterDictionary;
 use shared::bxdfs::HairBxDF;
 use shared::core::material::Material;
 use shared::core::spectrum::Spectrum;
 use shared::core::texture::SpectrumType;
 use shared::materials::complex::*;
-// use shared::spectra::SampledWavelengths;
 use shared::textures::SpectrumConstantTexture;
-// use shared::utils::Ptr;
 use anyhow::Result;
 use std::collections::HashMap;
 use std::sync::Arc;
diff --git a/src/shapes/bilinear.rs b/src/shapes/bilinear.rs
index 403a80d..dbbd1da 100644
--- a/src/shapes/bilinear.rs
+++ b/src/shapes/bilinear.rs
@@ -1,14 +1,14 @@
 use crate::core::image::{Image, ImageIO};
-use crate::core::shape::{ALL_BILINEAR_MESHES, CreateShape};
+use crate::core::shape::{CreateShape, ALL_BILINEAR_MESHES};
 use crate::core::texture::FloatTexture;
 use crate::shapes::mesh::BilinearPatchMesh;
 use crate::utils::sampling::PiecewiseConstant2D;
-use crate::utils::{Arena, FileLoc, ParameterDictionary};
-use anyhow::{Result, anyhow};
+use crate::{Arena, FileLoc, ParameterDictionary};
+use anyhow::{anyhow, Result};
 use log::warn;
 use shared::core::shape::Shape;
 use shared::shapes::BilinearPatchShape;
-use shared::utils::{Ptr, Transform};
+use shared::{Ptr, Transform};
 use std::collections::HashMap;
 use std::path::Path;
 use std::sync::Arc;
diff --git a/src/shapes/curves.rs b/src/shapes/curves.rs
index 7fed644..61ed919 100644
--- a/src/shapes/curves.rs
+++ b/src/shapes/curves.rs
@@ -1,20 +1,17 @@
-use crate::Arena;
 use crate::core::shape::CreateShape;
 use crate::core::texture::FloatTexture;
-use crate::utils::{FileLoc, ParameterDictionary};
-use anyhow::{Result, anyhow};
-use shared::Float;
+use crate::{Arena, FileLoc, ParameterDictionary};
+use anyhow::{anyhow, Result};
+use log::warn;
 use shared::core::geometry::{Normal3f, Point3f};
 use shared::core::shape::Shape;
 use shared::shapes::{CurveCommon, CurveShape, CurveType};
-use shared::utils::Transform;
 use shared::utils::math::lerp;
 use shared::utils::splines::{
     cubic_bspline_to_bezier, elevate_quadratic_bezier_to_cubic, quadratic_bspline_to_bezier,
 };
-use shared::Ptr;
-
-use log::warn;
+use shared::Float;
+use shared::{Ptr, Transform};
 use std::collections::HashMap;
 use std::sync::Arc;
 
@@ -199,7 +196,7 @@ impl CreateShape for CurveShape {
                 curve_type,
                 seg_normals.expect("Could not determine normals to curve segments"),
                 split_depth.try_into().unwrap(),
-                arena
+                arena,
             );
 
             curves.extend(new_curves);
diff --git a/src/shapes/cylinder.rs b/src/shapes/cylinder.rs
index 7ab7463..3461dc0 100644
--- a/src/shapes/cylinder.rs
+++ b/src/shapes/cylinder.rs
@@ -1,13 +1,12 @@
 use crate::core::shape::CreateShape;
 use crate::core::texture::FloatTexture;
-use crate::utils::{Arena, FileLoc, ParameterDictionary};
+use crate::{Arena, FileLoc, ParameterDictionary};
 use anyhow::Result;
 use shared::core::shape::Shape;
 use shared::shapes::CylinderShape;
-use shared::utils::Transform;
 use std::collections::HashMap;
 use std::sync::Arc;
-use shared::Ptr;
+use shared::{Transform, Ptr};
 
 impl CreateShape for CylinderShape {
     fn create(
diff --git a/src/shapes/disk.rs b/src/shapes/disk.rs
index 8b5eee6..1caa6e0 100644
--- a/src/shapes/disk.rs
+++ b/src/shapes/disk.rs
@@ -1,13 +1,12 @@
 use crate::core::shape::CreateShape;
 use crate::core::texture::FloatTexture;
-use crate::utils::{Arena, FileLoc, ParameterDictionary};
+use crate::{Arena, FileLoc, ParameterDictionary};
 use anyhow::Result;
 use shared::core::shape::Shape;
 use shared::shapes::DiskShape;
-use shared::utils::Transform;
+use shared::{Ptr, Transform};
 use std::collections::HashMap;
 use std::sync::Arc;
-use shared::Ptr;
 
 impl CreateShape for DiskShape {
     fn create(
diff --git a/src/shapes/mesh.rs b/src/shapes/mesh.rs
index ee62212..edb4d85 100644
--- a/src/shapes/mesh.rs
+++ b/src/shapes/mesh.rs
@@ -1,16 +1,14 @@
-// use crate::Arena;
 use crate::utils::sampling::PiecewiseConstant2D;
+use crate::{Arena, Device, DeviceRepr};
 use anyhow::{bail, Context, Result as AnyResult};
 use ply_rs::parser::Parser;
 use ply_rs::ply::{DefaultElement, Property};
 use shared::core::geometry::{Normal3f, Point2f, Point3f, Vector3f, VectorLike};
-use shared::utils::mesh::{DeviceBilinearPatchMesh, DeviceTriangleMesh};
-use shared::utils::{Ptr, Transform};
+use shared::shapes::mesh::{BilinearPatchMesh, TriangleMesh};
+use shared::{Ptr, Transform};
 use std::fs::File;
-use std::ops::Deref;
 use std::path::Path;
 
-/// Intermediate mesh from PLY
 #[derive(Debug, Clone, Default)]
 pub struct TriQuadMesh {
     pub p: Vec<Point3f>,
@@ -21,235 +19,7 @@ pub struct TriQuadMesh {
     pub quad_indices: Vec<i32>,
 }
 
-#[derive(DeviceRepr)]
-#[device(name = "DeviceTriangleMesh")]
-pub struct TriangleMeshStorage {
-    pub vertex_indices: Vec<i32>,    // → Ptr<i32> + len (always present)
-    pub p: Vec<Point3f>,             // → Ptr<Point3f> + len (always present)
-    pub n: Vec<Normal3f>,            // → Ptr<Normal3f> + len (empty → null Ptr, len 0)
-    pub s: Vec<Vector3f>,            // → Ptr<Vector3f> + len
-    pub uv: Vec<Point2f>,            // → Ptr<Point2f> + len
-    pub face_indices: Vec<i32>,      // → Ptr<i32> + len
-}
-
-
-#[derive(Debug)]
-pub(crate) struct BilinearMeshStorage {
-    pub vertex_indices: Vec<i32>,
-    pub p: Vec<Point3f>,
-    pub n: Vec<Normal3f>,
-    pub uv: Vec<Point2f>,
-    pub image_distribution: Option<PiecewiseConstant2D>,
-}
-
-#[derive(Debug)]
-pub struct TriangleMesh {
-    pub storage: Box<TriangleMeshStorage>,
-    pub device: DeviceTriangleMesh,
-}
-
-#[derive(Debug)]
-pub struct BilinearPatchMesh {
-    pub storage: Box<BilinearMeshStorage>,
-    pub device: DeviceBilinearPatchMesh,
-}
-
-impl Deref for TriangleMesh {
-    type Target = DeviceTriangleMesh;
-    #[inline]
-    fn deref(&self) -> &Self::Target {
-        &self.device
-    }
-}
-
-impl Deref for BilinearPatchMesh {
-    type Target = DeviceBilinearPatchMesh;
-    #[inline]
-    fn deref(&self) -> &Self::Target {
-        &self.device
-    }
-}
-
-impl TriangleMesh {
-    pub fn new(
-        render_from_object: &Transform,
-        reverse_orientation: bool,
-        vertex_indices: Vec<i32>,
-        mut p: Vec<Point3f>,
-        mut n: Vec<Normal3f>,
-        mut s: Vec<Vector3f>,
-        uv: Vec<Point2f>,
-        face_indices: Vec<i32>,
-    ) -> Self {
-        let n_triangles = vertex_indices.len() / 3;
-        let n_vertices = p.len();
-
-        // Transform positions to render space
-        for pt in p.iter_mut() {
-            *pt = render_from_object.apply_to_point(*pt);
-        }
-
-        // Transform and optionally flip normals
-        if !n.is_empty() {
-            assert_eq!(n_vertices, n.len(), "Normal count mismatch");
-            for nn in n.iter_mut() {
-                *nn = render_from_object.apply_to_normal(*nn);
-                if reverse_orientation {
-                    *nn = -*nn;
-                }
-            }
-        }
-
-        // Transform tangents
-        if !s.is_empty() {
-            assert_eq!(n_vertices, s.len(), "Tangent count mismatch");
-            for ss in s.iter_mut() {
-                *ss = render_from_object.apply_to_vector(*ss);
-            }
-        }
-
-        // Validate UVs
-        if !uv.is_empty() {
-            assert_eq!(n_vertices, uv.len(), "UV count mismatch");
-        }
-
-        // Validate face indices
-        if !face_indices.is_empty() {
-            assert_eq!(n_triangles, face_indices.len(), "Face index count mismatch");
-        }
-
-        let transform_swaps_handedness = render_from_object.swaps_handedness();
-
-        let storage = Box::new(TriangleMeshStorage {
-            vertex_indices,
-            p,
-            n,
-            s,
-            uv,
-            face_indices,
-        });
-
-        // Build device struct with pointers into storage
-        let device = DeviceTriangleMesh {
-            n_triangles: n_triangles as u32,
-            n_vertices: n_vertices as u32,
-            vertex_indices: storage.vertex_indices.as_ptr().into(),
-            p: storage.p.as_ptr().into(),
-            n: if storage.n.is_empty() {
-                Ptr::null()
-            } else {
-                storage.n.as_ptr().into()
-            },
-            s: if storage.s.is_empty() {
-                Ptr::null()
-            } else {
-                storage.s.as_ptr().into()
-            },
-            uv: if storage.uv.is_empty() {
-                Ptr::null()
-            } else {
-                storage.uv.as_ptr().into()
-            },
-            face_indices: if storage.face_indices.is_empty() {
-                Ptr::null()
-            } else {
-                storage.face_indices.as_ptr().into()
-            },
-            reverse_orientation,
-            transform_swaps_handedness,
-        };
-
-        Self { storage, device }
-    }
-
-    pub fn positions(&self) -> &[Point3f] {
-        &self.storage.p
-    }
-
-    pub fn indices(&self) -> &[i32] {
-        &self.storage.vertex_indices
-    }
-
-    pub fn normals(&self) -> &[Normal3f] {
-        &self.storage.n
-    }
-
-    pub fn uvs(&self) -> &[Point2f] {
-        &self.storage.uv
-    }
-}
-
-impl BilinearPatchMesh {
-    pub fn new(
-        render_from_object: &Transform,
-        reverse_orientation: bool,
-        vertex_indices: Vec<i32>,
-        mut p: Vec<Point3f>,
-        mut n: Vec<Normal3f>,
-        uv: Vec<Point2f>,
-        image_distribution: Option<PiecewiseConstant2D>,
-    ) -> Self {
-        let n_patches = vertex_indices.len() / 4;
-        let n_vertices = p.len();
-
-        for pt in p.iter_mut() {
-            *pt = render_from_object.apply_to_point(*pt);
-        }
-
-        if !n.is_empty() {
-            assert_eq!(n_vertices, n.len(), "Normal count mismatch");
-            for nn in n.iter_mut() {
-                *nn = render_from_object.apply_to_normal(*nn);
-                if reverse_orientation {
-                    *nn = -*nn;
-                }
-            }
-        }
-
-        if !uv.is_empty() {
-            assert_eq!(n_vertices, uv.len(), "UV count mismatch");
-        }
-
-        let transform_swaps_handedness = render_from_object.swaps_handedness();
-
-        let storage = Box::new(BilinearMeshStorage {
-            vertex_indices,
-            p,
-            n,
-            uv,
-            image_distribution,
-        });
-
-        let device = DeviceBilinearPatchMesh {
-            n_patches: n_patches as u32,
-            n_vertices: n_vertices as u32,
-            vertex_indices: storage.vertex_indices.as_ptr().into(),
-            p: storage.p.as_ptr().into(),
-            n: if storage.n.is_empty() {
-                Ptr::null()
-            } else {
-                storage.n.as_ptr().into()
-            },
-            uv: if storage.uv.is_empty() {
-                Ptr::null()
-            } else {
-                storage.uv.as_ptr().into()
-            },
-            reverse_orientation,
-            transform_swaps_handedness,
-            image_distribution: storage
-                .image_distribution
-                .as_ref()
-                .map(|d| Ptr::from(&d.device))
-                .unwrap_or(Ptr::null()),
-        };
-
-        Self { storage, device }
-    }
-}
-
 // PLY Helper Functions
-
 fn get_float(elem: &DefaultElement, key: &str) -> AnyResult<f32> {
     match elem.get(key) {
         Some(Property::Float(v)) => Ok(*v),
@@ -290,14 +60,12 @@ fn get_list_uint(elem: &DefaultElement, key: &str) -> AnyResult<Vec<u32>> {
     }
 }
 
-// TriQuadMesh Implementation
-
 impl TriQuadMesh {
     pub fn read_ply<P: AsRef<Path>>(filename: P) -> AnyResult<Self> {
         let path = filename.as_ref();
         let filename_display = path.display().to_string();
 
-        let mut f = File::open(path)
+        let f = File::open(path)
             .with_context(|| format!("Couldn't open PLY file \"{}\"", filename_display))?;
 
         let p = Parser::<DefaultElement>::new();
@@ -313,7 +81,6 @@ impl TriQuadMesh {
 
         let mut mesh = TriQuadMesh::default();
 
-        // Parse vertices
         if let Some(vertices) = ply.payload.get("vertex") {
             if vertices.is_empty() {
                 bail!("{}: PLY file has no vertices", filename_display);
@@ -360,7 +127,6 @@ impl TriQuadMesh {
             bail!("{}: PLY file has no vertex elements", filename_display);
         }
 
-        // Parse faces
         if let Some(faces) = ply.payload.get("face") {
             mesh.tri_indices.reserve(faces.len() * 3);
             mesh.quad_indices.reserve(faces.len() * 4);
@@ -390,7 +156,6 @@ impl TriQuadMesh {
             bail!("{}: PLY file has no face elements", filename_display);
         }
 
-        // Validate indices
         let vertex_count = mesh.p.len() as i32;
         for &idx in &mesh.tri_indices {
             if idx >= vertex_count {
@@ -421,18 +186,13 @@ impl TriQuadMesh {
             return;
         }
 
-        // Each quad becomes 2 triangles
         self.tri_indices.reserve(self.quad_indices.len() / 4 * 6);
 
         for quad in self.quad_indices.chunks_exact(4) {
             let (v0, v1, v2, v3) = (quad[0], quad[1], quad[2], quad[3]);
-
-            // Triangle 1: v0, v1, v2
             self.tri_indices.push(v0);
             self.tri_indices.push(v1);
             self.tri_indices.push(v2);
-
-            // Triangle 2: v0, v2, v3
             self.tri_indices.push(v0);
             self.tri_indices.push(v2);
             self.tri_indices.push(v3);
@@ -442,10 +202,8 @@ impl TriQuadMesh {
     }
 
     pub fn compute_normals(&mut self) {
-        // Initialize normals to zero
         self.n = vec![Normal3f::new(0.0, 0.0, 0.0); self.p.len()];
 
-        // Accumulate face normals for triangles
         for tri in self.tri_indices.chunks_exact(3) {
             let (i0, i1, i2) = (tri[0] as usize, tri[1] as usize, tri[2] as usize);
 
@@ -457,13 +215,11 @@ impl TriQuadMesh {
             let v20 = p2 - p0;
             let face_normal = v10.cross(v20);
 
-            // Accumulate (will normalize later)
             self.n[i0] = self.n[i0] + Normal3f::from(face_normal);
             self.n[i1] = self.n[i1] + Normal3f::from(face_normal);
             self.n[i2] = self.n[i2] + Normal3f::from(face_normal);
         }
 
-        // Accumulate face normals for quads
         for quad in self.quad_indices.chunks_exact(4) {
             let indices: Vec<usize> = quad.iter().map(|&i| i as usize).collect();
 
@@ -480,7 +236,6 @@ impl TriQuadMesh {
             }
         }
 
-        // Normalize all normals
         for normal in &mut self.n {
             let len_sq = normal.norm_squared();
             if len_sq > 0.0 {
@@ -489,7 +244,6 @@ impl TriQuadMesh {
         }
     }
 
-    /// Convert to a TriangleMesh, consuming self
     pub fn into_triangle_mesh(
         mut self,
         render_from_object: &Transform,
@@ -507,15 +261,13 @@ impl TriQuadMesh {
             self.tri_indices,
             self.p,
             self.n,
-            Vec::new(), // s (tangents)
+            Vec::new(),
             self.uv,
             self.face_indices,
         )
     }
 }
 
-// Create from PLY directly
-
 impl TriangleMesh {
     pub fn from_ply<P: AsRef<Path>>(
         filename: P,
diff --git a/src/shapes/sphere.rs b/src/shapes/sphere.rs
index a88d2b4..56c222e 100644
--- a/src/shapes/sphere.rs
+++ b/src/shapes/sphere.rs
@@ -1,13 +1,12 @@
 use crate::core::shape::CreateShape;
 use crate::core::texture::FloatTexture;
-use crate::utils::{Arena, FileLoc, ParameterDictionary};
+use crate::{Arena, FileLoc, ParameterDictionary};
 use anyhow::Result;
 use shared::core::shape::Shape;
 use shared::shapes::SphereShape;
-use shared::utils::Transform;
+use shared::{Ptr, Transform};
 use std::collections::HashMap;
 use std::sync::Arc;
-use shared::Ptr;
 
 impl CreateShape for SphereShape {
     fn create(
diff --git a/src/shapes/triangle.rs b/src/shapes/triangle.rs
index ab387af..fba4bfd 100644
--- a/src/shapes/triangle.rs
+++ b/src/shapes/triangle.rs
@@ -1,12 +1,12 @@
-use crate::core::shape::{ALL_TRIANGLE_MESHES, CreateShape};
+use crate::core::shape::{CreateShape, ALL_TRIANGLE_MESHES};
 use crate::core::texture::FloatTexture;
 use crate::shapes::mesh::TriangleMesh;
-use crate::utils::{Arena, FileLoc, ParameterDictionary};
-use anyhow::{Result, bail};
+use crate::{Arena, FileLoc, ParameterDictionary};
+use anyhow::{bail, Result};
 use log::warn;
 use shared::core::shape::Shape;
 use shared::shapes::TriangleShape;
-use shared::utils::{Ptr, Transform};
+use shared::{Ptr, Transform};
 use std::collections::HashMap;
 use std::sync::Arc;
 
diff --git a/src/textures/mix.rs b/src/textures/mix.rs
index 8cdc5da..93ab749 100644
--- a/src/textures/mix.rs
+++ b/src/textures/mix.rs
@@ -1,20 +1,28 @@
-use crate::Arena;
 use crate::core::texture::{
     CreateSpectrumTexture, FloatTexture, FloatTextureTrait, SpectrumTexture, SpectrumTextureTrait,
 };
 use crate::utils::{FileLoc, TextureParameterDictionary};
+use crate::{Arena, Device, DeviceRepr};
 use anyhow::Result;
-use shared::Float;
 use shared::core::geometry::{Vector3f, VectorLike};
 use shared::core::texture::{SpectrumType, TextureEvalContext};
 use shared::spectra::{SampledSpectrum, SampledWavelengths};
+use shared::textures::{
+    GPUFloatDirectionMixTexture, GPUFloatMixTexture, GPUSpectrumDirectionMixTexture,
+    GPUSpectrumMixTexture,
+};
 use shared::utils::Transform;
+use shared::Float;
 use std::sync::Arc;
 
-#[derive(Debug, Clone)]
+#[derive(Clone, Debug, Device)]
+#[device(name = "GPUFloatMixTexture")]
 pub struct FloatMixTexture {
+    #[device(upload)]
     pub tex1: Arc<FloatTexture>,
+    #[device(upload)]
     pub tex2: Arc<FloatTexture>,
+    #[device(upload)]
     pub amount: Arc<FloatTexture>,
 }
 
@@ -56,9 +64,12 @@ impl FloatTextureTrait for FloatMixTexture {
     }
 }
 
-#[derive(Debug, Clone)]
+#[derive(Clone, Debug, Device)]
+#[device(name = "GPUFloatDirectionMixTexture")]
 pub struct FloatDirectionMixTexture {
+    #[device(upload)]
     pub tex1: Arc<FloatTexture>,
+    #[device(upload)]
     pub tex2: Arc<FloatTexture>,
     pub dir: Vector3f,
 }
@@ -89,10 +100,14 @@ impl FloatTextureTrait for FloatDirectionMixTexture {
     }
 }
 
-#[derive(Debug, Clone)]
+#[derive(Clone, Debug, Device)]
+#[device(name = "GPUSpectrumMixTexture")]
 pub struct SpectrumMixTexture {
+    #[device(upload)]
     pub tex1: Arc<SpectrumTexture>,
+    #[device(upload)]
     pub tex2: Arc<SpectrumTexture>,
+    #[device(upload)]
     pub amount: Arc<FloatTexture>,
 }
 
@@ -113,9 +128,12 @@ impl SpectrumTextureTrait for SpectrumMixTexture {
     }
 }
 
-#[derive(Debug, Clone)]
+#[derive(Clone, Debug, Device)]
+#[device(name = "GPUSpectrumDirectionMixTexture")]
 pub struct SpectrumDirectionMixTexture {
+    #[device(upload)]
     pub tex1: Arc<SpectrumTexture>,
+    #[device(upload)]
     pub tex2: Arc<SpectrumTexture>,
     pub dir: Vector3f,
 }
diff --git a/src/textures/scaled.rs b/src/textures/scaled.rs
index 239f085..19c909c 100644
--- a/src/textures/scaled.rs
+++ b/src/textures/scaled.rs
@@ -1,17 +1,21 @@
-use crate::Arena;
 use crate::core::texture::{CreateSpectrumTexture, FloatTexture, SpectrumTexture};
 use crate::core::texture::{FloatTextureTrait, SpectrumTextureTrait};
 use crate::utils::{FileLoc, TextureParameterDictionary};
+use crate::{Arena, Device, DeviceRepr};
 use anyhow::Result;
-use shared::Float;
 use shared::core::texture::{SpectrumType, TextureEvalContext};
 use shared::spectra::{SampledSpectrum, SampledWavelengths};
+use shared::textures::{GPUFloatScaledTexture, GPUSpectrumScaledTexture};
 use shared::utils::Transform;
+use shared::Float;
 use std::sync::Arc;
 
-#[derive(Debug, Clone)]
+#[derive(Clone, Debug, Device)]
+#[device(name = "GPUFloatScaledTexture")]
 pub struct FloatScaledTexture {
+    #[device(upload)]
     pub tex: Arc<FloatTexture>,
+    #[device(upload)]
     pub scale: Arc<FloatTexture>,
 }
 
@@ -44,7 +48,6 @@ impl FloatScaledTexture {
             std::mem::swap(&mut tex, &mut scale);
         }
         std::mem::swap(&mut tex, &mut scale);
-        // arena.alloc(FloatScaledTexture::new(tex, scale));
         Ok(FloatTexture::Scaled(FloatScaledTexture::new(
             tex.clone(),
             scale.clone(),
@@ -62,9 +65,12 @@ impl FloatTextureTrait for FloatScaledTexture {
     }
 }
 
-#[derive(Clone, Debug)]
+#[derive(Clone, Debug, Device)]
+#[device(name = "GPUSpectrumScaledTexture")]
 pub struct SpectrumScaledTexture {
+    #[device(upload)]
     pub tex: Arc<SpectrumTexture>,
+    #[device(upload)]
     pub scale: Arc<FloatTexture>,
 }
 
diff --git a/src/utils/arena.rs b/src/utils/arena.rs
index 485a183..426c9d0 100644
--- a/src/utils/arena.rs
+++ b/src/utils/arena.rs
@@ -1,7 +1,5 @@
 use crate::utils::backend::GpuAllocator;
-use shared::core::shape::Shape;
-use shared::core::material::Material;
-use shared::core::spectrum::Spectrum;
+use crate::utils::mipmap::MIPMap;
 use parking_lot::Mutex;
 use shared::Ptr;
 use std::alloc::Layout;
@@ -173,378 +171,3 @@ impl<A: GpuAllocator + Default> Default for Arena<A> {
         Self::new(A::default())
     }
 }
-
-pub trait DeviceRepr {
-    /// The `#[repr(C)] Copy` device-side struct.
-    type Target: Copy;
-
-    /// Upload into the arena and return the device struct by value.
-    /// Use this when embedding the result inline in another device struct.
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target;
-
-    /// Upload into the arena and return a Ptr to the device struct.
-    /// This is the common entry point — allocates the Target in the arena.
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let value = self.upload_value(arena);
-        arena.alloc(value)
-    }
-}
-
-impl<T: DeviceRepr> DeviceRepr for Option<T> {
-    type Target = T::Target;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
-        match self {
-            Some(val) => val.upload_value(arena),
-            None => panic!("Cannot upload_value on None — use upload() which returns Ptr::null()"),
-        }
-    }
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        match self {
-            Some(val) => val.upload(arena),
-            None => Ptr::null(),
-        }
-    }
-}
-
-impl<T: DeviceRepr> DeviceRepr for std::sync::Arc<T> {
-    type Target = T::Target;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
-        (**self).upload_value(arena)
-    }
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        (**self).upload(arena)
-    }
-}
-
-impl<T: DeviceRepr> DeviceRepr for Box<T> {
-    type Target = T::Target;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
-        (**self).upload_value(arena)
-    }
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        (**self).upload(arena)
-    }
-}
-
-impl DeviceRepr for Shape {
-    type Target = Shape;
-    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> Shape {
-        self.clone()
-    }
-}
-
-impl DeviceRepr for Light {
-    type Target = Light;
-    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> Light {
-        self.clone()
-    }
-}
-
-impl DeviceRepr for Spectrum {
-    type Target = Spectrum;
-    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> Spectrum {
-        self.clone()
-    }
-}
-
-impl DeviceRepr for Material {
-    type Target = Material;
-    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> Material {
-        self.clone()
-    }
-}
-
-// =============================================================================
-// Image → DeviceImage
-// =============================================================================
-
-impl DeviceRepr for Image {
-    type Target = DeviceImage;
-    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> DeviceImage {
-        *self.device()
-    }
-}
-
-// =============================================================================
-// DenselySampledSpectrumBuffer → DenselySampledSpectrum
-// =============================================================================
-
-impl DeviceRepr for DenselySampledSpectrumBuffer {
-    type Target = DenselySampledSpectrum;
-    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> DenselySampledSpectrum {
-        self.device()
-    }
-}
-
-// =============================================================================
-// RGBToSpectrumTable — re-uploads Ptr fields into arena
-// =============================================================================
-
-impl DeviceRepr for RGBToSpectrumTable {
-    type Target = RGBToSpectrumTable;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> RGBToSpectrumTable {
-        let n_nodes = self.n_nodes as usize;
-
-        // Safety: these Ptrs point into static or previously-uploaded data;
-        // we're copying the contents into the arena for a new lifetime.
-        let z_slice = unsafe { from_raw_parts(self.z_nodes.as_raw(), n_nodes) };
-        let (z_ptr, _) = arena.alloc_slice(z_slice);
-
-        let n_coeffs = 3 * n_nodes.pow(3);
-        let coeffs_slice = unsafe { from_raw_parts(self.coeffs.as_raw(), n_coeffs) };
-        let (c_ptr, _) = arena.alloc_slice(coeffs_slice);
-
-        RGBToSpectrumTable {
-            z_nodes: z_ptr,
-            coeffs: c_ptr,
-            n_nodes: self.n_nodes,
-        }
-    }
-}
-
-// =============================================================================
-// RGBColorSpace — nested upload of spectrum table
-// =============================================================================
-
-impl DeviceRepr for RGBColorSpace {
-    type Target = RGBColorSpace;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> RGBColorSpace {
-        let table_ptr = self.rgb_to_spectrum_table.upload(arena);
-
-        RGBColorSpace {
-            r: self.r,
-            g: self.g,
-            b: self.b,
-            w: self.w,
-            illuminant: self.illuminant.clone(),
-            rgb_to_spectrum_table: table_ptr,
-            xyz_from_rgb: self.xyz_from_rgb,
-            rgb_from_xyz: self.rgb_from_xyz,
-        }
-    }
-}
-
-// =============================================================================
-// DeviceStandardColorSpaces — composition of color space uploads
-// =============================================================================
-
-impl DeviceRepr for DeviceStandardColorSpaces {
-    type Target = DeviceStandardColorSpaces;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> DeviceStandardColorSpaces {
-        DeviceStandardColorSpaces {
-            srgb: self.srgb.upload(arena),
-            dci_p3: self.dci_p3.upload(arena),
-            rec2020: self.rec2020.upload(arena),
-            aces2065_1: self.aces2065_1.upload(arena),
-        }
-    }
-}
-
-// =============================================================================
-// TriangleMesh → DeviceTriangleMesh
-// =============================================================================
-// TriangleMesh should own all its fields directly — no .device field.
-// The host struct holds Vec arrays + scalar metadata. derive(Device) handles it:
-//
-//   #[derive(Device)]
-//   #[device(name = "DeviceTriangleMesh")]
-//   pub struct TriangleMesh {
-//       pub vertex_indices: Vec<i32>,
-//       pub p: Vec<Point3f>,
-//       pub n: Vec<Normal3f>,
-//       pub s: Vec<Vector3f>,
-//       pub uv: Vec<Point2f>,
-//       pub face_indices: Vec<i32>,
-//       pub n_triangles: u32,
-//       pub reverse_orientation: bool,
-//       pub transform_swaps_handedness: bool,
-//   }
-//
-// Until the mesh struct is refactored to remove .device, here's a manual
-// impl that lists every field. This is the MIGRATION TARGET — once TriangleMesh
-// drops its .device field and puts scalars at the top level, switch to derive(Device).
-
-impl DeviceRepr for TriangleMesh {
-    type Target = DeviceTriangleMesh;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> DeviceTriangleMesh {
-        let s = &self.storage;
-
-        let (vertex_indices_ptr, _) = arena.alloc_slice(&s.vertex_indices);
-        let (p_ptr, _) = arena.alloc_slice(&s.p);
-        let (n_ptr, _) = arena.alloc_slice(&s.n);
-        let (s_ptr, _) = arena.alloc_slice(&s.s);
-        let (uv_ptr, _) = arena.alloc_slice(&s.uv);
-        let (face_indices_ptr, _) = arena.alloc_slice(&s.face_indices);
-
-        DeviceTriangleMesh {
-            vertex_indices: vertex_indices_ptr,
-            p: p_ptr,
-            n: n_ptr,
-            s: s_ptr,
-            uv: uv_ptr,
-            face_indices: face_indices_ptr,
-            n_triangles: self.n_triangles,
-            reverse_orientation: self.reverse_orientation,
-            transform_swaps_handedness: self.transform_swaps_handedness,
-        }
-    }
-}
-
-// =============================================================================
-// BilinearPatchMesh → DeviceBilinearPatchMesh
-// =============================================================================
-// Same as TriangleMesh: once the host struct is refactored to own scalars
-// directly (no .device field), this switches to derive(Device).
-
-impl DeviceRepr for BilinearPatchMesh {
-    type Target = DeviceBilinearPatchMesh;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> DeviceBilinearPatchMesh {
-        let s = &self.storage;
-
-        let (vertex_indices_ptr, _) = arena.alloc_slice(&s.vertex_indices);
-        let (p_ptr, _) = arena.alloc_slice(&s.p);
-        let (n_ptr, _) = arena.alloc_slice(&s.n);
-        let (uv_ptr, _) = arena.alloc_slice(&s.uv);
-
-        let image_dist_ptr = s.image_distribution.upload(arena);
-
-        DeviceBilinearPatchMesh {
-            vertex_indices: vertex_indices_ptr,
-            p: p_ptr,
-            n: n_ptr,
-            uv: uv_ptr,
-            image_distribution: image_dist_ptr,
-            n_patches: self.n_patches,
-            reverse_orientation: self.reverse_orientation,
-            transform_swaps_handedness: self.transform_swaps_handedness,
-        }
-    }
-}
-
-// =============================================================================
-// SpectrumTexture → GPUSpectrumTexture
-// =============================================================================
-
-impl DeviceRepr for SpectrumTexture {
-    type Target = GPUSpectrumTexture;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> GPUSpectrumTexture {
-        match self {
-            SpectrumTexture::Constant(tex) => GPUSpectrumTexture::Constant(tex.clone()),
-            SpectrumTexture::Checkerboard(tex) => GPUSpectrumTexture::Checkerboard(tex.clone()),
-            SpectrumTexture::Dots(tex) => GPUSpectrumTexture::Dots(tex.clone()),
-            SpectrumTexture::Image(tex) => {
-                let tex_obj = arena.get_texture_object(&tex.base.mipmap);
-                GPUSpectrumTexture::Image(GPUSpectrumImageTexture {
-                    mapping: tex.base.mapping,
-                    tex_obj,
-                    scale: tex.base.scale,
-                    invert: tex.base.invert,
-                    is_single_channel: tex.base.mipmap.is_single_channel(),
-                    color_space: tex
-                        .base
-                        .mipmap
-                        .color_space
-                        .clone()
-                        .unwrap_or_else(crate::spectra::default_colorspace),
-                    spectrum_type: tex.spectrum_type,
-                })
-            }
-            SpectrumTexture::Bilerp(tex) => GPUSpectrumTexture::Bilerp(tex.clone()),
-            SpectrumTexture::Scaled(tex) => {
-                let child_ptr = tex.tex.upload(arena);
-                let scale_ptr = tex.scale.upload(arena);
-                GPUSpectrumTexture::Scaled(GPUSpectrumScaledTexture {
-                    tex: child_ptr,
-                    scale: scale_ptr,
-                })
-            }
-            SpectrumTexture::Marble(tex) => GPUSpectrumTexture::Marble(tex.clone()),
-            SpectrumTexture::Mix(tex) => {
-                let tex1_ptr = tex.tex1.upload(arena);
-                let tex2_ptr = tex.tex2.upload(arena);
-                let amount_ptr = tex.amount.upload(arena);
-                GPUSpectrumTexture::Mix(GPUSpectrumMixTexture {
-                    tex1: tex1_ptr,
-                    tex2: tex2_ptr,
-                    amount: amount_ptr,
-                })
-            }
-            SpectrumTexture::DirectionMix(tex) => {
-                let tex1_ptr = tex.tex1.upload(arena);
-                let tex2_ptr = tex.tex2.upload(arena);
-                GPUSpectrumTexture::DirectionMix(GPUSpectrumDirectionMixTexture {
-                    tex1: tex1_ptr,
-                    tex2: tex2_ptr,
-                    dir: tex.dir,
-                })
-            }
-        }
-    }
-}
-
-// =============================================================================
-// FloatTexture → GPUFloatTexture
-// =============================================================================
-
-impl DeviceRepr for FloatTexture {
-    type Target = GPUFloatTexture;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> GPUFloatTexture {
-        match self {
-            FloatTexture::Constant(tex) => GPUFloatTexture::Constant(tex.clone()),
-            FloatTexture::Checkerboard(tex) => GPUFloatTexture::Checkerboard(tex.clone()),
-            FloatTexture::Dots(tex) => GPUFloatTexture::Dots(tex.clone()),
-            FloatTexture::FBm(tex) => GPUFloatTexture::FBm(tex.clone()),
-            FloatTexture::Windy(tex) => GPUFloatTexture::Windy(tex.clone()),
-            FloatTexture::Wrinkled(tex) => GPUFloatTexture::Wrinkled(tex.clone()),
-            FloatTexture::Scaled(tex) => {
-                let child_ptr = tex.tex.upload(arena);
-                let scale_ptr = tex.scale.upload(arena);
-                GPUFloatTexture::Scaled(GPUFloatScaledTexture {
-                    tex: child_ptr,
-                    scale: scale_ptr,
-                })
-            }
-            FloatTexture::Mix(tex) => {
-                let tex1_ptr = tex.tex1.upload(arena);
-                let tex2_ptr = tex.tex2.upload(arena);
-                let amount_ptr = tex.amount.upload(arena);
-                GPUFloatTexture::Mix(GPUFloatMixTexture {
-                    tex1: tex1_ptr,
-                    tex2: tex2_ptr,
-                    amount: amount_ptr,
-                })
-            }
-            FloatTexture::DirectionMix(tex) => {
-                let tex1_ptr = tex.tex1.upload(arena);
-                let tex2_ptr = tex.tex2.upload(arena);
-                GPUFloatTexture::DirectionMix(GPUFloatDirectionMixTexture {
-                    tex1: tex1_ptr,
-                    tex2: tex2_ptr,
-                    dir: tex.dir,
-                })
-            }
-            FloatTexture::Image(tex) => {
-                GPUFloatTexture::Image(GPUFloatImageTexture {
-                    mapping: tex.base.mapping,
-                    tex_obj: tex.base.mipmap.texture_object(),
-                    scale: tex.base.scale,
-                    invert: tex.base.invert,
-                })
-            }
-            FloatTexture::Bilerp(tex) => GPUFloatTexture::Bilerp(tex.clone()),
-        }
-    }
-}
diff --git a/src/utils/backend.rs b/src/utils/backend.rs
index 9c44412..3535f1e 100644
--- a/src/utils/backend.rs
+++ b/src/utils/backend.rs
@@ -20,12 +20,12 @@ impl GpuAllocator for SystemAllocator {
         if layout.size() == 0 {
             return layout.align() as *mut u8;
         }
-        std::alloc::alloc(layout)
+        unsafe { std::alloc::alloc(layout) }
     }
 
     unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
         if layout.size() > 0 {
-            std::alloc::dealloc(ptr, layout);
+            unsafe { std::alloc::dealloc(ptr, layout) };
         }
     }
 }
diff --git a/src/utils/containers.rs b/src/utils/containers.rs
index 128b0f2..5a767e4 100644
--- a/src/utils/containers.rs
+++ b/src/utils/containers.rs
@@ -1,9 +1,9 @@
+use crate::{Arena, DeviceRepr};
 use parking_lot::Mutex;
 use shared::core::geometry::{Bounds2i, Point2i};
 use shared::utils::containers::DeviceArray2D;
 use std::collections::HashSet;
 use std::hash::Hash;
-use std::ops::{Deref, DerefMut};
 use std::sync::Arc;
 
 pub struct InternCache<T> {
@@ -32,63 +32,3 @@ where
         new_item
     }
 }
-
-#[derive(Debug, Clone)]
-#[derive(DeviceRepr)]
-pub struct Array2D<T> {
-    pub device: DeviceArray2D<T>,
-    pub values: Vec<T>,
-}
-
-impl<T> Deref for Array2D<T> {
-    type Target = DeviceArray2D<T>;
-    fn deref(&self) -> &Self::Target {
-        &self.device
-    }
-}
-
-impl<T> DerefMut for Array2D<T> {
-    fn deref_mut(&mut self) -> &mut Self::Target {
-        &mut self.device
-    }
-}
-
-impl<T> Array2D<T> {
-    pub fn as_slice(&self) -> &[T] {
-        &self.values
-    }
-
-    fn from_vec(extent: Bounds2i, mut values: Vec<T>) -> Self {
-        let width = extent.p_max.x() - extent.p_min.x();
-        let device = DeviceArray2D {
-            extent,
-            values: values.as_mut_ptr().into(),
-            stride: width,
-        };
-        Self { device, values }
-    }
-
-    pub fn device(&self) -> &DeviceArray2D<T> {
-        &self.device
-    }
-}
-
-impl<T: Default + Clone> Array2D<T> {
-    pub fn new(extent: Bounds2i) -> Self {
-        let n = extent.area() as usize;
-        let storage = vec![T::default(); n];
-        Self::from_vec(extent, storage)
-    }
-
-    pub fn new_dims(nx: i32, ny: i32) -> Self {
-        let extent = Bounds2i::from_points(Point2i::new(0, 0), Point2i::new(nx, ny));
-        Self::new(extent)
-    }
-
-    pub fn new_filled(nx: i32, ny: i32, val: T) -> Self {
-        let extent = Bounds2i::from_points(Point2i::new(0, 0), Point2i::new(nx, ny));
-        let n = (nx * ny) as usize;
-        let storage = vec![val; n];
-        Self::from_vec(extent, storage)
-    }
-}
diff --git a/src/utils/math.rs b/src/utils/math.rs
index a1ea46e..500f919 100644
--- a/src/utils/math.rs
+++ b/src/utils/math.rs
@@ -1,9 +1,7 @@
 use half::f16;
-use shared::Float;
 use shared::utils::hash::hash_buffer;
-use shared::utils::math::{
-    DeviceDigitPermutation, PRIMES, f16_to_f32_software, permutation_element,
-};
+use shared::utils::math::{permutation_element, DeviceDigitPermutation, PRIMES};
+use shared::Float;
 
 #[inline(always)]
 pub fn f16_to_f32(bits: u16) -> f32 {
diff --git a/src/utils/mod.rs b/src/utils/mod.rs
index 18ba826..e294c8b 100644
--- a/src/utils/mod.rs
+++ b/src/utils/mod.rs
@@ -11,14 +11,16 @@ pub mod parameters;
 pub mod parser;
 pub mod sampling;
 pub mod strings;
+pub mod upload;
 
-pub use arena::DeviceRepr;
+pub use device_derive::Device;
 pub use error::FileLoc;
 pub use file::{read_float_file, resolve_filename};
 pub use parameters::{
     ParameterDictionary, ParsedParameter, ParsedParameterVector, TextureParameterDictionary,
 };
 pub use strings::*;
+pub use upload::DeviceRepr;
 
 #[cfg(feature = "vulkan")]
 pub type Arena = arena::Arena<backend::vulkan::VulkanAllocator>;
@@ -28,516 +30,3 @@ pub type Arena = arena::Arena<backend::cuda::CudaAllocator>;
 
 #[cfg(not(any(feature = "cuda", feature = "vulkan")))]
 pub type Arena = arena::Arena<backend::SystemAllocator>;
-
-/// # Enum variant attributes
-///
-/// | Attribute | Effect |
-/// |-----------|--------|
-/// | *(none)* | Inner type has `DeviceRepr`; auto-call `upload_value` |
-/// | `#[device(clone)]` | Same type on both sides, just clone |
-/// | `#[device(custom = "method")]` | You provide `fn method(inner: &T, arena) -> DeviceT` |
-/// | `#[device(variant_type = "T")]` | Override the device-side variant's inner type |
-///
-/// # Container attribute
-///
-/// `#[device(name = "DeviceFoo")]` — override the generated type name (default: `Device{Name}`).
-#[proc_macro_derive(Device, attributes(device))]
-pub fn derive_device(input: TokenStream) -> TokenStream {
-    let input = parse_macro_input!(input as DeriveInput);
-    match derive_impl(input) {
-        Ok(tokens) => tokens.into(),
-        Err(e) => e.to_compile_error().into(),
-    }
-}
-
-fn derive_impl(input: DeriveInput) -> syn::Result<TokenStream2> {
-    match &input.data {
-        Data::Struct(_) => derive_struct(input),
-        Data::Enum(_) => derive_enum(input),
-        Data::Union(_) => Err(syn::Error::new_spanned(
-            &input.ident,
-            "Device derive does not support unions",
-        )),
-    }
-}
-
-// Struct derivation
-
-fn derive_struct(input: DeriveInput) -> syn::Result<TokenStream2> {
-    let host_name = &input.ident;
-    let vis = &input.vis;
-    let device_name = get_device_name(&input.attrs, host_name)?;
-
-    let fields = match &input.data {
-        Data::Struct(s) => match &s.fields {
-            Fields::Named(named) => &named.named,
-            _ => {
-                return Err(syn::Error::new_spanned(
-                    host_name,
-                    "Device derive only supports structs with named fields",
-                ))
-            }
-        },
-        _ => unreachable!(),
-    };
-
-    let mut device_fields = Vec::new();
-    let mut upload_stmts = Vec::new();
-    let mut device_field_inits = Vec::new();
-    let mut spread_expr: Option<Expr> = None;
-
-    for field in fields {
-        let field_name = field.ident.as_ref().unwrap();
-        let attrs = parse_field_attrs(&field.attrs)?;
-
-        if attrs.skip {
-            continue;
-        }
-
-        if let Some(ref expr_str) = attrs.spread {
-            spread_expr = Some(syn::parse_str(expr_str).map_err(|e| {
-                syn::Error::new_spanned(field, format!("invalid device(spread): {}", e))
-            })?);
-            continue;
-        }
-
-        if let Some(expr_str) = &attrs.expr {
-            let expr: Expr = syn::parse_str(expr_str).map_err(|e| {
-                syn::Error::new_spanned(field, format!("invalid device(expr): {}", e))
-            })?;
-            let ty = &field.ty;
-            device_fields.push(quote! { pub #field_name: #ty });
-            upload_stmts.push(quote! { let #field_name = #expr; });
-            device_field_inits.push(quote! { #field_name });
-            continue;
-        }
-
-        match classify_type(&field.ty) {
-            FieldClass::VecCopy(inner_ty) => {
-                let len_name = format_ident!("{}_len", field_name);
-                device_fields.push(quote! { pub #field_name: Ptr<#inner_ty> });
-                device_fields.push(quote! { pub #len_name: usize });
-                upload_stmts.push(quote! {
-                    let (#field_name, #len_name) = arena.alloc_slice(&self.#field_name);
-                });
-                device_field_inits.push(quote! { #field_name });
-                device_field_inits.push(quote! { #len_name });
-            }
-            FieldClass::VecUploadable(inner_ty) => {
-                let len_name = format_ident!("{}_len", field_name);
-                device_fields.push(quote! {
-                    pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
-                });
-                device_fields.push(quote! { pub #len_name: usize });
-                upload_stmts.push(quote! {
-                    let __up: Vec<<#inner_ty as DeviceRepr>::Target> = self.#field_name
-                        .iter()
-                        .map(|item| DeviceRepr::upload_value(item, arena))
-                        .collect();
-                    let (#field_name, #len_name) = arena.alloc_slice(&__up);
-                });
-                device_field_inits.push(quote! { #field_name });
-                device_field_inits.push(quote! { #len_name });
-            }
-            FieldClass::Option(inner_ty) => {
-                device_fields.push(quote! {
-                    pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
-                });
-                upload_stmts.push(quote! {
-                    let #field_name = match &self.#field_name {
-                        Some(val) => DeviceRepr::upload(val, arena),
-                        None => Ptr::null(),
-                    };
-                });
-                device_field_inits.push(quote! { #field_name });
-            }
-            FieldClass::Arc(inner_ty) => {
-                if attrs.flatten {
-                    device_fields.push(quote! {
-                        pub #field_name: <#inner_ty as DeviceRepr>::Target
-                    });
-                    upload_stmts.push(quote! {
-                        let #field_name = DeviceRepr::upload_value(&*self.#field_name, arena);
-                    });
-                } else {
-                    device_fields.push(quote! {
-                        pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
-                    });
-                    upload_stmts.push(quote! {
-                        let #field_name = DeviceRepr::upload(&*self.#field_name, arena);
-                    });
-                }
-                device_field_inits.push(quote! { #field_name });
-            }
-            FieldClass::Plain => {
-                let ty = &field.ty;
-                if attrs.copy_upload {
-                    device_fields.push(quote! { pub #field_name: #ty });
-                    upload_stmts.push(quote! {
-                        let #field_name = self.#field_name.clone();
-                    });
-                } else if attrs.flatten {
-                    device_fields.push(quote! {
-                        pub #field_name: <#ty as DeviceRepr>::Target
-                    });
-                    upload_stmts.push(quote! {
-                        let #field_name = DeviceRepr::upload_value(&self.#field_name, arena);
-                    });
-                } else if attrs.upload {
-                    device_fields.push(quote! {
-                        pub #field_name: Ptr<<#ty as DeviceRepr>::Target>
-                    });
-                    upload_stmts.push(quote! {
-                        let #field_name = DeviceRepr::upload(&self.#field_name, arena);
-                    });
-                } else {
-                    device_fields.push(quote! { pub #field_name: #ty });
-                    upload_stmts.push(quote! {
-                        let #field_name = self.#field_name;
-                    });
-                }
-                device_field_inits.push(quote! { #field_name });
-            }
-        }
-    }
-
-    let constructor = if let Some(spread) = spread_expr {
-        quote! {
-            #device_name {
-                #(#device_field_inits,)*
-                ..#spread
-            }
-        }
-    } else {
-        quote! {
-            #device_name {
-                #(#device_field_inits,)*
-            }
-        }
-    };
-
-    Ok(quote! {
-        #[repr(C)]
-        #[derive(Debug, Copy, Clone)]
-        #vis struct #device_name {
-            #(#device_fields,)*
-        }
-
-        unsafe impl Send for #device_name {}
-        unsafe impl Sync for #device_name {}
-
-        impl DeviceRepr for #host_name {
-            type Target = #device_name;
-
-            fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
-                #(#upload_stmts)*
-                #constructor
-            }
-        }
-    })
-}
-
-// Enum derivation
-fn derive_enum(input: DeriveInput) -> syn::Result<TokenStream2> {
-    let host_name = &input.ident;
-    let vis = &input.vis;
-    let device_name = get_device_name(&input.attrs, host_name)?;
-
-    let variants = match &input.data {
-        Data::Enum(e) => &e.variants,
-        _ => unreachable!(),
-    };
-
-    let mut device_variants = Vec::new();
-    let mut match_arms = Vec::new();
-
-    for variant in variants {
-        let var_name = &variant.ident;
-        let var_attrs = parse_variant_attrs(&variant.attrs)?;
-        let inner_ty = get_variant_inner_type(variant)?;
-
-        // Determine the device-side inner type for this variant
-        let device_inner: Type = if let Some(ref ty_str) = var_attrs.variant_type {
-            syn::parse_str(ty_str).map_err(|e| {
-                syn::Error::new_spanned(variant, format!("invalid variant_type: {}", e))
-            })?
-        } else if var_attrs.clone_variant {
-            // clone: same type on both sides
-            inner_ty.clone()
-        } else {
-            // auto-upload: use DeviceRepr::Target
-            syn::parse_str(&format!("<{} as DeviceRepr>::Target", quote!(#inner_ty))).map_err(
-                |e| {
-                    syn::Error::new_spanned(variant, format!("cannot construct Target type: {}", e))
-                },
-            )?
-        };
-
-        device_variants.push(quote! { #var_name(#device_inner) });
-
-        if var_attrs.clone_variant {
-            match_arms.push(quote! {
-                #host_name::#var_name(inner) => #device_name::#var_name(inner.clone())
-            });
-        } else if let Some(ref method) = var_attrs.custom {
-            let method_ident = format_ident!("{}", method);
-            match_arms.push(quote! {
-                #host_name::#var_name(inner) => {
-                    #device_name::#var_name(Self::#method_ident(inner, arena))
-                }
-            });
-        } else {
-            // Default: inner implements DeviceRepr
-            match_arms.push(quote! {
-                #host_name::#var_name(inner) => {
-                    #device_name::#var_name(DeviceRepr::upload_value(inner, arena))
-                }
-            });
-        }
-    }
-
-    Ok(quote! {
-        #[repr(C)]
-        #[derive(Debug, Copy, Clone)]
-        #vis enum #device_name {
-            #(#device_variants,)*
-        }
-
-        unsafe impl Send for #device_name {}
-        unsafe impl Sync for #device_name {}
-
-        impl DeviceRepr for #host_name {
-            type Target = #device_name;
-
-            fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
-                match self {
-                    #(#match_arms,)*
-                }
-            }
-        }
-    })
-}
-
-fn get_variant_inner_type(variant: &Variant) -> syn::Result<Type> {
-    match &variant.fields {
-        Fields::Unnamed(fields) if fields.unnamed.len() == 1 => {
-            Ok(fields.unnamed.first().unwrap().ty.clone())
-        }
-        Fields::Unit => Err(syn::Error::new_spanned(
-            variant,
-            "Device derive: enum variants must have exactly one field, e.g. Variant(Type)",
-        )),
-        _ => Err(syn::Error::new_spanned(
-            variant,
-            "Device derive: only single-field tuple variants supported, e.g. Variant(Type)",
-        )),
-    }
-}
-
-// Attribute parsing for variants
-struct VariantAttrs {
-    clone_variant: bool,
-    custom: Option<String>,
-    variant_type: Option<String>,
-}
-
-fn parse_variant_attrs(attrs: &[Attribute]) -> syn::Result<VariantAttrs> {
-    let mut result = VariantAttrs {
-        clone_variant: false,
-        custom: None,
-        variant_type: None,
-    };
-
-    for attr in attrs {
-        if !attr.path().is_ident("device") {
-            continue;
-        }
-        attr.parse_nested_meta(|meta| {
-            if meta.path.is_ident("clone") {
-                result.clone_variant = true;
-                Ok(())
-            } else if meta.path.is_ident("custom") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    result.custom = Some(s.value());
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else if meta.path.is_ident("variant_type") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    result.variant_type = Some(s.value());
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else {
-                Err(meta.error("unknown device variant attribute"))
-            }
-        })?;
-    }
-
-    Ok(result)
-}
-
-// Attribute parsing for fields
-struct FieldAttrs {
-    skip: bool,
-    expr: Option<String>,
-    copy_upload: bool,
-    flatten: bool,
-    upload: bool,
-    spread: Option<String>,
-}
-
-fn parse_field_attrs(attrs: &[Attribute]) -> syn::Result<FieldAttrs> {
-    let mut result = FieldAttrs {
-        skip: false,
-        expr: None,
-        copy_upload: false,
-        flatten: false,
-        upload: false,
-        spread: None,
-    };
-
-    for attr in attrs {
-        if !attr.path().is_ident("device") {
-            continue;
-        }
-        attr.parse_nested_meta(|meta| {
-            if meta.path.is_ident("skip") {
-                result.skip = true;
-                Ok(())
-            } else if meta.path.is_ident("expr") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    result.expr = Some(s.value());
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else if meta.path.is_ident("copy_upload") {
-                result.copy_upload = true;
-                Ok(())
-            } else if meta.path.is_ident("flatten") {
-                result.flatten = true;
-                Ok(())
-            } else if meta.path.is_ident("upload") {
-                result.upload = true;
-                Ok(())
-            } else if meta.path.is_ident("spread") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    result.spread = Some(s.value());
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else {
-                Err(meta.error("unknown device attribute"))
-            }
-        })?;
-    }
-
-    Ok(result)
-}
-
-// Container-level name attribute
-fn get_device_name(attrs: &[Attribute], host_name: &Ident) -> syn::Result<Ident> {
-    for attr in attrs {
-        if !attr.path().is_ident("device") {
-            continue;
-        }
-        let mut name = None;
-        attr.parse_nested_meta(|meta| {
-            if meta.path.is_ident("name") {
-                let value = meta.value()?;
-                let lit: Lit = value.parse()?;
-                if let Lit::Str(s) = lit {
-                    name = Some(format_ident!("{}", s.value()));
-                    Ok(())
-                } else {
-                    Err(meta.error("expected string literal"))
-                }
-            } else {
-                Ok(())
-            }
-        })?;
-        if let Some(n) = name {
-            return Ok(n);
-        }
-    }
-    Ok(format_ident!("Device{}", host_name))
-}
-
-// Type classification
-enum FieldClass {
-    VecCopy(Type),
-    VecUploadable(Type),
-    Option(Type),
-    Arc(Type),
-    Plain,
-}
-
-fn classify_type(ty: &Type) -> FieldClass {
-    if let Some(inner) = extract_generic_arg(ty, "Vec") {
-        if is_copy_primitive(&inner) {
-            FieldClass::VecCopy(inner)
-        } else {
-            FieldClass::VecUploadable(inner)
-        }
-    } else if let Some(inner) = extract_generic_arg(ty, "Option") {
-        FieldClass::Option(inner)
-    } else if let Some(inner) = extract_generic_arg(ty, "Arc") {
-        FieldClass::Arc(inner)
-    } else {
-        FieldClass::Plain
-    }
-}
-
-fn extract_generic_arg(ty: &Type, wrapper: &str) -> Option<Type> {
-    if let Type::Path(type_path) = ty {
-        let seg = type_path.path.segments.last()?;
-        if seg.ident != wrapper {
-            return None;
-        }
-        if let PathArguments::AngleBracketed(args) = &seg.arguments {
-            if let Some(GenericArgument::Type(inner)) = args.args.first() {
-                return Some(inner.clone());
-            }
-        }
-    }
-    None
-}
-
-fn is_copy_primitive(ty: &Type) -> bool {
-    if let Type::Path(type_path) = ty {
-        if let Some(seg) = type_path.path.segments.last() {
-            let name = seg.ident.to_string();
-            return matches!(
-                name.as_str(),
-                "f32"
-                    | "f64"
-                    | "u8"
-                    | "u16"
-                    | "u32"
-                    | "u64"
-                    | "i8"
-                    | "i16"
-                    | "i32"
-                    | "i64"
-                    | "usize"
-                    | "isize"
-                    | "bool"
-                    | "Float"
-            );
-        }
-    }
-    false
-}
-
diff --git a/src/utils/sampling.rs b/src/utils/sampling.rs
index 37c87db..bfe026f 100644
--- a/src/utils/sampling.rs
+++ b/src/utils/sampling.rs
@@ -1,78 +1,35 @@
 use crate::core::image::Image;
-use crate::utils::arena::Arena;
-use crate::utils::backend::GpuAllocator;
 use crate::utils::containers::Array2D;
+use crate::{Arena, DeviceRepr};
 use shared::core::geometry::{Bounds2f, Point2i, Vector2f, Vector2i};
 use shared::utils::sampling::{
-    AliasTable, Bin, DevicePiecewiseConstant1D, DevicePiecewiseConstant2D, DeviceSummedAreaTable,
-    DeviceWindowedPiecewiseConstant2D, PiecewiseLinear2D,
+    AliasTable, Bin, PiecewiseConstant1D, PiecewiseConstant2D,
+    SummedAreaTable, WindowedPiecewiseConstant2D, PiecewiseLinear2D,
 };
 use shared::utils::{gpu_array_from_fn, Ptr};
 use shared::Float;
 use std::sync::Arc;
 
 #[derive(Debug, Clone)]
-pub struct PiecewiseConstant1D {
+pub struct HostPiecewiseConstant1D {
     func: Vec<Float>,
     cdf: Vec<Float>,
     pub min: Float,
     pub max: Float,
 }
 
-impl PiecewiseConstant1D {
-    pub fn new(f: &[Float]) -> Self {
-        Self::new_with_bounds(f.to_vec(), 0.0, 1.0)
-    }
-
-    pub fn from_func<F>(f: F, min: Float, max: Float, n: usize) -> Self
-    where
-        F: Fn(Float) -> Float,
-    {
-        let delta = (max - min) / n as Float;
-        let values: Vec<Float> = (0..n)
-            .map(|i| f(min + (i as Float + 0.5) * delta))
-            .collect();
-        Self::new_with_bounds(values, min, max)
-    }
-
-    pub fn new_with_bounds(f: Vec<Float>, min: Float, max: Float) -> Self {
-        let n = f.len();
-        let mut cdf = Vec::with_capacity(n + 1);
-        cdf.push(0.0);
-
-        let delta = (max - min) / n as Float;
-        for i in 0..n {
-            cdf.push(cdf[i] + f[i] * delta);
-        }
-
-        let func_integral = cdf[n];
-        if func_integral > 0.0 {
-            for c in &mut cdf {
-                *c /= func_integral;
-            }
-        }
-
-        Self { func: f, cdf, min, max }
-    }
-
-    // Accessors
+impl HostPiecewiseConstant1D {
     pub fn n(&self) -> usize { self.func.len() }
     pub fn func(&self) -> &[Float] { &self.func }
     pub fn cdf(&self) -> &[Float] { &self.cdf }
 
     pub fn integral(&self) -> Float {
-        // func_integral is the un-normalized sum. After normalization cdf[n] == 1.0,
-        // so we reconstruct from the last CDF entry before normalization.
-        // But since we normalized in-place, we need to store it. Let's compute it.
         let n = self.func.len();
         let delta = (self.max - self.min) / n as Float;
         self.func.iter().sum::<Float>() * delta
     }
 
-    /// Host-side sampling (for scene construction, not rendering).
-    /// During rendering, use the device struct via arena-uploaded Ptrs.
     pub fn sample_host(&self, u: Float) -> (Float, Float, usize) {
-        let n = self.func.len();
         let offset = self.find_interval_host(u);
         let cdf_offset = self.cdf[offset];
         let cdf_next = self.cdf[offset + 1];
@@ -81,6 +38,7 @@ impl PiecewiseConstant1D {
         } else {
             0.0
         };
+        let n = self.func.len();
         let delta = (self.max - self.min) / n as Float;
         let x = self.min + (offset as Float + du) * delta;
         let func_integral = self.integral();
@@ -110,41 +68,21 @@ impl PiecewiseConstant1D {
     }
 }
 
-#[derive(DeviceRepr)]
-#[device(name = "DevicePiecewiseConstant1D")]
-pub struct PiecewiseConstant1D {
-    pub func: Vec<Float>,
-    pub cdf: Vec<Float>,
-    pub min: Float,
-    pub max: Float,
-    pub n: u32,
-    #[device(expr = "self.integral()")]
-    pub func_integral: Float,
-}
-
-#[derive(Clone, Debug, DeviceRepr)]
-#[device(name = "DevicePiecewiseConstant2D")]
+#[derive(Debug, Clone)]
 pub struct PiecewiseConstant2D {
     pub conditionals: Vec<PiecewiseConstant1D>,
-    #[device(flatten)]
     pub marginal: PiecewiseConstant1D,
-    pub n_u: u32,
-    pub n_v: u32,
+    pub n_u: usize,
+    pub n_v: usize,
 }
 
-
 impl PiecewiseConstant2D {
     pub fn new(data: &Array2D<Float>) -> Self {
         Self::new_with_bounds(data, Bounds2f::unit())
     }
 
     pub fn new_with_bounds(data: &Array2D<Float>, domain: Bounds2f) -> Self {
-        Self::from_slice(
-            data.as_slice(),
-            data.x_size(),
-            data.y_size(),
-            domain,
-        )
+        Self::from_slice(data.as_slice(), data.x_size(), data.y_size(), domain)
     }
 
     pub fn from_slice(data: &[Float], n_u: usize, n_v: usize, domain: Bounds2f) -> Self {
@@ -155,20 +93,14 @@ impl PiecewiseConstant2D {
 
         for v in 0..n_v {
             let row = data[v * n_u..(v + 1) * n_u].to_vec();
-            let conditional = PiecewiseConstant1D::new_with_bounds(
-                row,
-                domain.p_min.x(),
-                domain.p_max.x(),
-            );
+            let conditional =
+                PiecewiseConstant1D::new_with_bounds(row, domain.p_min.x(), domain.p_max.x());
             marginal_func.push(conditional.integral());
             conditionals.push(conditional);
         }
 
-        let marginal = PiecewiseConstant1D::new_with_bounds(
-            marginal_func,
-            domain.p_min.y(),
-            domain.p_max.y(),
-        );
+        let marginal =
+            PiecewiseConstant1D::new_with_bounds(marginal_func, domain.p_min.y(), domain.p_max.y());
 
         Self { conditionals, marginal, n_u, n_v }
     }
@@ -181,7 +113,9 @@ impl PiecewiseConstant2D {
         let mut data = Vec::with_capacity(n_u * n_v);
         for v in 0..n_v {
             for u in 0..n_u {
-                data.push(image.get_channels(Point2i::new(u as i32, v as i32)).average());
+                data.push(
+                    image.get_channels(Point2i::new(u as i32, v as i32)).average(),
+                );
             }
         }
 
@@ -193,179 +127,22 @@ impl PiecewiseConstant2D {
     }
 }
 
-struct PiecewiseLinear2DStorage<const N: usize> {
-    data: Vec<Float>,
-    marginal_cdf: Vec<Float>,
-    conditional_cdf: Vec<Float>,
-    param_values: [Vec<Float>; N],
-}
+impl DeviceRepr for PiecewiseConstant2D {
+    type Target = DevicePiecewiseConstant2D;
 
-pub struct PiecewiseLinear2DHost<const N: usize> {
-    size: Vector2i,
-    inv_patch_size: Vector2f,
-    param_size: [u32; N],
-    param_strides: [u32; N],
-    storage: Arc<PiecewiseLinear2DStorage<N>>,
-}
+    fn upload_value(&self, arena: &Arena) -> DevicePiecewiseConstant2D {
+        let uploaded: Vec<DevicePiecewiseConstant1D> = self
+            .conditionals
+            .iter()
+            .map(|c| c.upload_value(arena))
+            .collect();
+        let (conditionals_ptr, _) = arena.alloc_slice(&uploaded);
 
-
-impl<const N: usize> PiecewiseLinear2DHost<N> {
-    pub fn new(
-        data: &[Float],
-        x_size: i32,
-        y_size: i32,
-        param_res: [usize; N],
-        param_values: [&[Float]; N],
-        normalize: bool,
-        build_cdf: bool,
-    ) -> Self {
-        if build_cdf && !normalize {
-            panic!("PiecewiseLinear2D::new: build_cdf implies normalize=true");
-        }
-
-        let size = Vector2i::new(x_size, y_size);
-        let inv_patch_size = Vector2f::new(1. / (x_size - 1) as Float, 1. / (y_size - 1) as Float);
-
-        let mut param_size = [0u32; N];
-        let mut param_strides = [0u32; N];
-        let param_values = gpu_array_from_fn(|i| param_values[i].to_vec());
-
-        let mut slices: u32 = 1;
-        for i in (0..N).rev() {
-            if param_res[i] < 1 {
-                panic!("PiecewiseLinear2D::new: parameter resolution must be >= 1!");
-            }
-            param_size[i] = param_res[i] as u32;
-            param_strides[i] = if param_res[i] > 1 { slices } else { 0 };
-            slices *= param_size[i];
-        }
-
-        let n_values = (x_size * y_size) as usize;
-        let mut new_data = vec![0.0; slices as usize * n_values];
-        let mut marginal_cdf = if build_cdf {
-            vec![0.0; slices as usize * y_size as usize]
-        } else {
-            Vec::new()
-        };
-        let mut conditional_cdf = if build_cdf {
-            vec![0.0; slices as usize * n_values]
-        } else {
-            Vec::new()
-        };
-
-        let mut data_offset = 0;
-        for slice in 0..slices as usize {
-            let slice_offset = slice * n_values;
-            let current_data = &data[data_offset..data_offset + n_values];
-            let mut sum = 0.;
-
-            // Construct conditional CDF
-            if normalize {
-                for y in 0..(y_size - 1) {
-                    for x in 0..(x_size - 1) {
-                        let i = (y * x_size + x) as usize;
-                        let v00 = current_data[i] as f64;
-                        let v10 = current_data[i + 1] as f64;
-                        let v01 = current_data[i + x_size as usize] as f64;
-                        let v11 = current_data[i + 1 + x_size as usize] as f64;
-                        sum += 0.25 * (v00 + v10 + v01 + v11);
-                    }
-                }
-            }
-
-            let normalization = if normalize && sum > 0.0 {
-                1.0 / sum as Float
-            } else {
-                1.0
-            };
-            for k in 0..n_values {
-                new_data[slice_offset + k] = current_data[k] * normalization;
-            }
-
-            if build_cdf {
-                let marginal_slice_offset = slice * y_size as usize;
-                // Construct marginal CDF
-                for y in 0..y_size as usize {
-                    let mut cdf_sum = 0.0;
-                    let i_base = y * x_size as usize;
-                    conditional_cdf[slice_offset + i_base] = 0.0;
-                    for x in 0..(x_size - 1) as usize {
-                        let i = i_base + x;
-                        cdf_sum +=
-                            0.5 * (new_data[slice_offset + i] + new_data[slice_offset + i + 1]);
-                        conditional_cdf[slice_offset + i + 1] = cdf_sum;
-                    }
-                }
-                // Construct marginal CDF
-                marginal_cdf[marginal_slice_offset] = 0.0;
-                let mut marginal_sum = 0.0;
-                for y in 0..(y_size - 1) as usize {
-                    let cdf1 = conditional_cdf[slice_offset + (y + 1) * x_size as usize - 1];
-                    let cdf2 = conditional_cdf[slice_offset + (y + 2) * x_size as usize - 1];
-                    marginal_sum += 0.5 * (cdf1 + cdf2);
-                    marginal_cdf[marginal_slice_offset + y + 1] = marginal_sum;
-                }
-            }
-            data_offset += n_values;
-        }
-
-        let view = PiecewiseLinear2D {
-            size,
-            inv_patch_size,
-            param_size,
-            param_strides,
-            param_values: param_values.each_ref().map(|x| x.as_ptr().into()),
-            data: Ptr::null(),
-            marginal_cdf: marginal_cdf.as_ptr().into(),
-            conditional_cdf: conditional_cdf.as_ptr().into(),
-        };
-
-        let storage = Arc::new(PiecewiseLinear2DStorage {
-            data: new_data,
-            marginal_cdf,
-            conditional_cdf,
-            param_values,
-        });
-
-        let mut final_view = view;
-        final_view.data = storage.data.as_ptr().into();
-        final_view.marginal_cdf = storage.marginal_cdf.as_ptr().into();
-        final_view.conditional_cdf = storage.conditional_cdf.as_ptr().into();
-        for i in 0..N {
-            final_view.param_values[i] = storage.param_values[i].as_ptr().into();
-        }
-
-        Self {
-            view: final_view,
-            _storage: storage,
-        }
-    }
-}
-
-impl<const N: usize> DeviceRepr for PiecewiseLinear2DHost<N> {
-    type Target = PiecewiseLinear2D<N>;
-
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> PiecewiseLinear2D<N> {
-        let s = &self.storage;
-
-        let (data_ptr, _) = arena.alloc_slice(&s.data);
-        let (marginal_ptr, _) = arena.alloc_slice(&s.marginal_cdf);
-        let (conditional_ptr, _) = arena.alloc_slice(&s.conditional_cdf);
-
-        let param_ptrs: [Ptr<Float>; N] = std::array::from_fn(|i| {
-            let (ptr, _) = arena.alloc_slice(&s.param_values[i]);
-            ptr
-        });
-
-        PiecewiseLinear2D {
-            size: self.size,
-            inv_patch_size: self.inv_patch_size,
-            param_size: self.param_size,
-            param_strides: self.param_strides,
-            param_values: param_ptrs,
-            data: data_ptr,
-            marginal_cdf: marginal_ptr,
-            conditional_cdf: conditional_ptr,
+        DevicePiecewiseConstant2D {
+            conditionals: conditionals_ptr,
+            marginal: self.marginal.upload_value(arena),
+            n_u: self.n_u as u32,
+            n_v: self.n_v as u32,
         }
     }
 }
@@ -411,10 +188,8 @@ impl AliasTableHost {
         while !under.is_empty() && !over.is_empty() {
             let un = under.pop().unwrap();
             let ov = over.pop().unwrap();
-
             bins[un.index].q = un.p_hat as Float;
             bins[un.index].alias = ov.index as u32;
-
             let p_excess = un.p_hat + ov.p_hat - 1.0;
             if p_excess < 1.0 {
                 under.push(Outcome { p_hat: p_excess, index: ov.index });
@@ -442,22 +217,17 @@ impl AliasTableHost {
 impl DeviceRepr for AliasTableHost {
     type Target = AliasTable;
 
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> AliasTable {
+    fn upload_value(&self, arena: &Arena) -> AliasTable {
         if self.bins.is_empty() {
             return AliasTable { bins: Ptr::null(), size: 0 };
         }
         let (bins_ptr, _) = arena.alloc_slice(&self.bins);
-        AliasTable {
-            bins: bins_ptr,
-            size: self.bins.len() as u32,
-        }
+        AliasTable { bins: bins_ptr, size: self.bins.len() as u32 }
     }
 }
 
-#[derive(Clone, Debug, DeviceRepr)]
-#[device(name = "DeviceSummedAreaTable")]
+#[derive(Clone, Debug)]
 pub struct SummedAreaTable {
-    #[device(flatten)]
     sum: Array2D<f64>,
 }
 
@@ -488,12 +258,19 @@ impl SummedAreaTable {
     }
 }
 
-#[derive(Clone, Debug, DeviceRepr)]
-#[device(name = "DeviceWindowedPiecewiseConstant2D")]
+impl DeviceRepr for SummedAreaTable {
+    type Target = DeviceSummedAreaTable;
+
+    fn upload_value(&self, arena: &Arena) -> DeviceSummedAreaTable {
+        DeviceSummedAreaTable {
+            sum: self.sum.upload_value(arena),
+        }
+    }
+}
+
+#[derive(Clone, Debug)]
 pub struct WindowedPiecewiseConstant2D {
-    #[device(flatten)]
     sat: SummedAreaTable,
-    #[device(flatten)]
     func: Array2D<Float>,
 }
 
@@ -504,6 +281,17 @@ impl WindowedPiecewiseConstant2D {
     }
 }
 
+impl DeviceRepr for WindowedPiecewiseConstant2D {
+    type Target = DeviceWindowedPiecewiseConstant2D;
+
+    fn upload_value(&self, arena: &Arena) -> DeviceWindowedPiecewiseConstant2D {
+        DeviceWindowedPiecewiseConstant2D {
+            sat: self.sat.upload_value(arena),
+            func: self.func.upload_value(arena),
+        }
+    }
+}
+
 struct PiecewiseLinear2DStorage<const N: usize> {
     data: Vec<Float>,
     marginal_cdf: Vec<Float>,
@@ -534,10 +322,8 @@ impl<const N: usize> PiecewiseLinear2DHost<N> {
         }
 
         let size = Vector2i::new(x_size, y_size);
-        let inv_patch_size = Vector2f::new(
-            1.0 / (x_size - 1) as Float,
-            1.0 / (y_size - 1) as Float,
-        );
+        let inv_patch_size =
+            Vector2f::new(1.0 / (x_size - 1) as Float, 1.0 / (y_size - 1) as Float);
 
         let mut param_size = [0u32; N];
         let mut param_strides = [0u32; N];
@@ -600,18 +386,16 @@ impl<const N: usize> PiecewiseLinear2DHost<N> {
                     conditional_cdf[slice_offset + i_base] = 0.0;
                     for x in 0..(x_size - 1) as usize {
                         let i = i_base + x;
-                        cdf_sum += 0.5
-                            * (new_data[slice_offset + i] + new_data[slice_offset + i + 1]);
+                        cdf_sum +=
+                            0.5 * (new_data[slice_offset + i] + new_data[slice_offset + i + 1]);
                         conditional_cdf[slice_offset + i + 1] = cdf_sum;
                     }
                 }
                 marginal_cdf[marginal_slice_offset] = 0.0;
                 let mut marginal_sum = 0.0;
                 for y in 0..(y_size - 1) as usize {
-                    let cdf1 =
-                        conditional_cdf[slice_offset + (y + 1) * x_size as usize - 1];
-                    let cdf2 =
-                        conditional_cdf[slice_offset + (y + 2) * x_size as usize - 1];
+                    let cdf1 = conditional_cdf[slice_offset + (y + 1) * x_size as usize - 1];
+                    let cdf2 = conditional_cdf[slice_offset + (y + 2) * x_size as usize - 1];
                     marginal_sum += 0.5 * (cdf1 + cdf2);
                     marginal_cdf[marginal_slice_offset + y + 1] = marginal_sum;
                 }
@@ -626,20 +410,14 @@ impl<const N: usize> PiecewiseLinear2DHost<N> {
             param_values: owned_param_values,
         });
 
-        Self {
-            size,
-            inv_patch_size,
-            param_size,
-            param_strides,
-            storage,
-        }
+        Self { size, inv_patch_size, param_size, param_strides, storage }
     }
 }
 
 impl<const N: usize> DeviceRepr for PiecewiseLinear2DHost<N> {
     type Target = PiecewiseLinear2D<N>;
 
-    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> PiecewiseLinear2D<N> {
+    fn upload_value(&self, arena: &Arena) -> PiecewiseLinear2D<N> {
         let s = &self.storage;
 
         let (data_ptr, _) = arena.alloc_slice(&s.data);