Upload the learned demosaic's result, and blend grain back into it

DemosaicedImage::from_rgb_f32 takes the network's linear camera RGB and
stands it beside the classical source of the same photograph: the matrix,
profile tables and as-shot balance are that source's, the id is new, so
nothing downstream can tell which demosaic ran and every cache keyed on the
source sees a new one.

GrainBlend is the denoise's live control. It returns only the brightness of
the noise the network removed, taken after the as-shot balance and handed
back divided by it, so the grain is neutral in the finished picture; colour
speckle and demosaic false colour stay out. It writes a new source rather
than adding a term to the adjust shader: the blend depends on two images and
one number, a 20 MP pass is milliseconds, and a fresh source id is all the
adjust pass's caches need. The test reads it back: at 0 the network's
result, at 1 the same white-balanced step in every channel.
This commit is contained in:
2026-10-03 11:20:48 -04:00
parent d8304d7c82
commit 8ea3c3181a
5 changed files with 470 additions and 15 deletions
+89
View File
@@ -346,6 +346,95 @@ impl DemosaicedImage {
}
impl DemosaicedImage {
/// TRACES: FR-DEV-3g
/// The learned demosaic's output for the photograph `like` was
/// demosaiced from: `width × height` interleaved RGB, linear camera
/// space, normalised as the demosaic normalises — the same texture the
/// classical path made, with the noise gone (denoise.md §2).
///
/// Everything that describes the photograph rather than its pixels —
/// matrix, profile tables, as-shot balance — is `like`'s, so nothing
/// downstream can tell which demosaic ran. A new [`Self::id`], so every
/// cache keyed on the source sees a new source.
pub fn from_rgb_f32(
ctx: &GpuContext,
like: &DemosaicedImage,
width: u32,
height: u32,
rgb: &[f32],
) -> Result<Self, GpuError> {
let n = width as usize * height as usize;
if rgb.len() != n * 3 {
return Err(GpuError::TooLarge(format!(
"{} values for a {width}×{height} RGB image",
rgb.len()
)));
}
let limits = ctx.device.limits();
if width > limits.max_texture_dimension_2d || height > limits.max_texture_dimension_2d {
return Err(GpuError::TooLarge(format!(
"{width}×{height} exceeds the device limit of {}",
limits.max_texture_dimension_2d
)));
}
let mut half = vec![0u16; n * 4];
let one = f32_to_f16_bits(1.0);
let threads = std::thread::available_parallelism().map_or(1, |n| n.get());
let per = n.div_ceil(threads).max(1);
std::thread::scope(|scope| {
for (k, out) in half.chunks_mut(per * 4).enumerate() {
scope.spawn(move || {
for (i, texel) in out.chunks_mut(4).enumerate() {
let src = &rgb[(k * per + i) * 3..(k * per + i) * 3 + 3];
for c in 0..3 {
texel[c] = f32_to_f16_bits_unclamped(src[c]);
}
texel[3] = one;
}
});
}
});
let texture = ctx.device.create_texture_with_data(
&ctx.queue,
&wgpu::TextureDescriptor {
label: Some("learned-demosaic-source"),
size: wgpu::Extent3d {
width,
height,
depth_or_array_layers: 1,
},
mip_level_count: 1,
sample_count: 1,
dimension: wgpu::TextureDimension::D2,
format: Self::FORMAT,
usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_SRC,
view_formats: &[],
},
wgpu::util::TextureDataOrder::LayerMajor,
bytemuck::cast_slice(&half),
);
Ok(like.sibling(texture, width, height))
}
/// A new source standing for the same photograph as `self`: its
/// description kept, its pixels `texture`, a fresh id.
pub(crate) fn sibling(&self, texture: wgpu::Texture, width: u32, height: u32) -> Self {
let view = texture.create_view(&Default::default());
Self {
texture,
view,
width,
height,
color_matrix: self.color_matrix,
profile_tables: self.profile_tables.clone(),
as_shot_wb: self.as_shot_wb,
non_linear: self.non_linear,
id: next_image_id(),
frame: self.frame,
window: self.window,
}
}
/// TRACES: FR-MRG-3
/// A source that is already RGB in camera space: a linear DNG, which is
/// what a merge writes. No demosaic; the samples are normalised by the