diff --git a/crates/color/src/lib.rs b/crates/color/src/lib.rs
--- a/crates/color/src/lib.rs
+++ b/crates/color/src/lib.rs
@@ -20,7 +20,7 @@ pub use grade::GradeSpace;
 pub use log::LogCurve;
 pub use lut::{Lut, Lut1d, Lut3d, LutFormat};
 pub use spaces::{ColorPipeline, ColorSpace, Curve, Gamut, WorkingSpace};
-pub use transform::{DecodeTable, GpuCurve, GpuInputParams, InputTransform, OutputTransform, REFERENCE_WHITE_NITS, ToneMap};
+pub use transform::{DecodeTable, GpuCurve, GpuInputParams, GpuOutputParams, InputTransform, OutputTransform, REFERENCE_WHITE_NITS, ToneMap};
 
 /// YUV ↔ RGB matrix coefficients.
 #[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Hash, Serialize, Deserialize)]
diff --git a/crates/color/src/transform.rs b/crates/color/src/transform.rs
--- a/crates/color/src/transform.rs
+++ b/crates/color/src/transform.rs
@@ -389,6 +389,37 @@ impl OutputTransform {
             curve => l.map(|v| encode_channel(v as f64, curve, Range::Full).clamp(0.0, 1.0) as f32),
         }
     }
+
+    /// The transform as plain numbers for a shader, which runs [`apply`](Self::apply)'s stages
+    /// (BT.2390 tone mapping, the working → target matrix, desaturation into the target gamut) and
+    /// encodes the target's SDR curve itself. `None` for a target the shader does not encode (an
+    /// HDR export target), which stays on the CPU.
+    pub fn gpu_params(&self) -> Option<GpuOutputParams> {
+        if self.target.curve() != Curve::Sdr {
+            return None;
+        }
+        Some(GpuOutputParams {
+            target: self.target,
+            matrix: (!self.identity_matrix).then_some(self.matrix),
+            tone: self.tone.as_ref().map(|t| (t.src_peak_nits as f32, t.dst_peak_nits as f32)),
+            compress: self.compress,
+        })
+    }
+}
+
+/// [`OutputTransform`] as plain numbers for a shader ([`OutputTransform::gpu_params`]): the stages
+/// after the compositor — tone mapping, working → target matrix, gamut desaturation, in the order
+/// [`OutputTransform::apply`] runs them — with the target's SDR curve left to the shader's resolve.
+#[derive(Clone, Copy, Debug, PartialEq)]
+pub struct GpuOutputParams {
+    /// The target colour space (SDR: the shader's resolve is the sRGB curve).
+    pub target: ColorSpace,
+    /// Working → target primaries (row-major), when they differ.
+    pub matrix: Option<[[f32; 3]; 3]>,
+    /// BT.2390 tone mapping (source peak, target peak) in cd/m², on max(R,G,B) as a ratio.
+    pub tone: Option<(f32, f32)>,
+    /// Target luma weights to desaturate into the target gamut with ([`desaturate_into_gamut`]).
+    pub compress: Option<[f32; 3]>,
 }
 
 /// Convenience: a log colour space → display Rec. 709 (tone mapped), used for the built-in
@@ -550,4 +581,20 @@ mod tests {
         // plain pipeline is an identity
         assert!(OutputTransform::display(&ColorPipeline::REC709).is_identity());
     }
+
+    #[test]
+    fn output_gpu_params_cover_the_display_transform() {
+        // a plain pipeline: the shader's sRGB resolve is the whole transform
+        let plain = OutputTransform::display(&ColorPipeline::REC709).gpu_params().unwrap();
+        assert!(plain.matrix.is_none() && plain.tone.is_none() && plain.compress.is_none());
+        // an HDR working space on a SDR monitor: tone mapping, the working → 709 matrix and the
+        // gamut compression, the stages the resolve pass runs
+        let hdr = OutputTransform::display(&self::hdr(WorkingSpace::Rec2100Hlg)).gpu_params().unwrap();
+        assert_eq!(hdr.target, ColorSpace::Rec709);
+        assert!(hdr.matrix.is_some() && hdr.compress.is_some());
+        assert_eq!(hdr.tone, Some((HDR_PEAK_NITS as f32, REFERENCE_WHITE_NITS as f32)));
+        // an HDR export target has no shader encode yet: it stays on the CPU
+        let export = OutputTransform::new(&self::hdr(WorkingSpace::Rec2100Hlg), ColorSpace::Rec2100Pq);
+        assert!(export.gpu_params().is_none());
+    }
 }
diff --git a/crates/golden/tests/golden.rs b/crates/golden/tests/golden.rs
--- a/crates/golden/tests/golden.rs
+++ b/crates/golden/tests/golden.rs
@@ -775,6 +775,9 @@ fn gpu_matches_cpu_on_golden_scenes() {
         let cpu = render_cpu(&s);
         let plan = filmcraft_render::plan::plan_frame(&s.project, s.seq, t, RenderOptions::default(), &s.sources);
         let kind = if matches!(plan, filmcraft_render::plan::FramePlan::Layers { .. }) { "layers" } else { "cpu image" };
+        // the resolve pass runs the sequence's display transform, as the app sets it (an HDR
+        // working space composites in working linear and is tone mapped there)
+        c.set_display_transform(display_for(&s, &plan));
         c.composite(&plan);
         let (gw, gh, mut px) = c.read_output().expect("GPU readback");
         px.as_chunks_mut::<4>().0.iter_mut().for_each(|p| p[3] = 255);
@@ -793,6 +796,44 @@ fn gpu_matches_cpu_on_golden_scenes() {
     assert!(failures.is_empty(), "GPU differs from the CPU reference:\n{}", failures.join("\n"));
 }
 
+/// The resolve pass's display transform for `plan` of scene `s`, chosen as the app chooses it
+/// (`gpu_present`): none for a CPU-rendered frame, which is already in monitor space.
+fn display_for(s: &Scene, plan: &filmcraft_render::plan::FramePlan) -> Option<filmcraft_color::GpuOutputParams> {
+    let pipe = s.project.sequence(s.seq).map(|q| q.settings.color);
+    filmcraft_gpu::sequence_display(plan, pipe.and_then(|p| filmcraft_render::colorman::display_transform(&p).gpu_params()))
+}
+
+/// A frame of an HDR (Rec. 2100 PQ) sequence that the compositor cannot finish (a Wipe) comes
+/// back as one CPU image, already converted for the monitor: the GPU shows it as it is, without
+/// tone mapping it a second time.
+#[test]
+fn gpu_shows_an_hdr_cpu_frame_as_rendered() {
+    let Some((dev, q)) = device("gpu hdr cpu frame", false) else {
+        return;
+    };
+    let mut b = Builder::new(1);
+    b.p.sequence_mut(b.seq).unwrap().settings.color =
+        filmcraft_color::ColorPipeline { working: filmcraft_color::WorkingSpace::Rec2100Pq, ..filmcraft_color::ColorPipeline::REC709 };
+    let pq = ColorInfo { transfer: Transfer::Pq, primaries: Primaries::Bt2020, ..ColorInfo::SRGB_FULL };
+    let a = b.encoded(DemoScene::CityNight, ColorSpace::Rec2100Pq, 6.0, Some(pq), None);
+    let z = b.demo(DemoScene::Aurora);
+    let ca = b.place(0, a, 0, 24);
+    let cz = b.place(0, z, 24, 24);
+    b.transition(0, "wipe", ca, cz, 24, 12);
+    let s = b.at(24);
+    let t = RATE.tick_of(s.frame);
+    let plan = filmcraft_render::plan::plan_frame(&s.project, s.seq, t, RenderOptions::default(), &s.sources);
+    assert!(matches!(plan, filmcraft_render::plan::FramePlan::Image(_)), "a Wipe is rendered on the CPU");
+    let mut c = filmcraft_gpu::GpuCompositor::new(&dev, &q);
+    c.set_display_transform(display_for(&s, &plan));
+    c.composite(&plan);
+    let (gw, gh, mut px) = c.read_output().expect("GPU readback");
+    px.as_chunks_mut::<4>().0.iter_mut().for_each(|p| p[3] = 255);
+    let d = diff(&render_cpu(&s), &Rgba8::new(gw, gh, px)).unwrap();
+    let mean_rgb = d.mean_abs * 4.0 / 3.0;
+    assert!(d.p99 <= 6 && mean_rgb < 1.5, "HDR CPU frame on the GPU differs from the CPU: {d}");
+}
+
 /// Per-pixel max RGB difference (99th percentile, mean) over the pixels at least 1.5 px from every
 /// plan layer's quad outline: the CPU resampler fades a layer's edge over one more pixel than the
 /// GPU rasterises, and with blend modes those edge pixels are a visible share of a 320×180 frame.
diff --git a/crates/gpu/src/composite.wgsl b/crates/gpu/src/composite.wgsl
--- a/crates/gpu/src/composite.wgsl
+++ b/crates/gpu/src/composite.wgsl
@@ -394,6 +394,21 @@ fn fs_blend(in: VOut) -> @location(0) vec4<f32> {
 
 @group(0) @binding(0) var accum: texture_2d<f32>;
 
+// The resolve pass's display transform (`filmcraft_color::OutputTransform`, #965): the sequence's
+// working linear → monitor. Same stages as `OutputTransform::apply`, in the same order: BT.2390
+// tone mapping (from an HDR working space), the working → target gamut matrix, then desaturation
+// into the target gamut. `a.x` is 0 for a plain sequence (the accumulator is already monitor
+// space) and the sRGB encode below is the whole transform.
+struct O {
+    a: vec4<f32>,   // on, target curve (0 sdr), tone mapping on, matrix on
+    b: vec4<f32>,   // desaturation on, tone source peak cd/m², tone target peak cd/m², reference white cd/m²
+    m0: vec4<f32>,  // matrix row 0, desaturation luma weight 0
+    m1: vec4<f32>,  // matrix row 1, desaturation luma weight 1
+    m2: vec4<f32>,  // matrix row 2, desaturation luma weight 2
+};
+
+@group(0) @binding(1) var<uniform> disp: O;
+
 @vertex
 fn vs_full(@builtin(vertex_index) vi: u32) -> @builtin(position) vec4<f32> {
     let x = f32((vi << 1u) & 2u) * 2.0 - 1.0;
@@ -401,9 +416,44 @@ fn vs_full(@builtin(vertex_index) vi: u32) -> @builtin(position) vec4<f32> {
     return vec4(x, y, 0.0, 1.0);
 }
 
+// Working linear → monitor linear.
+fn display(rgb: vec3<f32>) -> vec3<f32> {
+    if disp.a.x < 0.5 { return rgb; }
+    var c = rgb;
+    if disp.a.z > 0.5 {
+        // `ToneMap::apply`: one scale from max(R,G,B), the EETF sampled in the PQ domain
+        let m = max(max(c.r, c.g), c.b);
+        if m > 0.0 {
+            let nits = m * disp.b.w;
+            var k = 1.0;
+            if nits >= disp.b.y {
+                // above the source peak: clip to the target peak
+                k = disp.b.z / disp.b.w / m;
+            } else if nits > 1e-6 {
+                k = eetf(nits, disp.b.y, disp.b.z) / nits;
+            }
+            c = c * k;
+        }
+    }
+    if disp.a.w > 0.5 { c = vec3(dot(disp.m0.xyz, c), dot(disp.m1.xyz, c), dot(disp.m2.xyz, c)); }
+    if disp.b.x > 0.5 {
+        // `desaturate_into_gamut`: out-of-gamut colours towards their luminance, in-gamut exactly
+        let mn = min(min(c.r, c.g), c.b);
+        if mn < 0.0 {
+            let y = dot(vec3(disp.m0.w, disp.m1.w, disp.m2.w), c);
+            if y <= 0.0 {
+                c = vec3(0.0);
+            } else {
+                c = max(y + (c - y) * (y / (y - mn)), vec3(0.0));
+            }
+        }
+    }
+    return c;
+}
+
 @fragment
 fn fs_full(@builtin(position) pos: vec4<f32>) -> @location(0) vec4<f32> {
-    let c = clamp(textureLoad(accum, vec2<i32>(pos.xy), 0).rgb, vec3(0.0), vec3(1.0));
+    let c = clamp(display(textureLoad(accum, vec2<i32>(pos.xy), 0).rgb), vec3(0.0), vec3(1.0));
     let lo = c * 12.92;
     let hi = 1.055 * pow(c, vec3(1.0 / 2.4)) - 0.055;
     return vec4(select(hi, lo, c <= vec3(0.0031308)), 1.0);
diff --git a/crates/gpu/src/fx_tests.rs b/crates/gpu/src/fx_tests.rs
--- a/crates/gpu/src/fx_tests.rs
+++ b/crates/gpu/src/fx_tests.rs
@@ -710,6 +710,44 @@ fn layers_with_effects_fall_back_to_the_cpu_without_a_stage() {
     }
 }
 
+/// A colour-managed layer (PQ media) with effects comes out the same with and without the GPU
+/// effect stage: the CPU fallback decodes it through the layer's input transform too.
+#[test]
+fn cpu_effect_fallback_keeps_the_input_transform() {
+    use filmcraft_color::{ColorInfo, ColorPipeline, ColorSpace, InputTransform, Primaries};
+    let Some((dev, q)) = device() else {
+        eprintln!("no GPU adapter; skipping");
+        return;
+    };
+    let (w, h) = (256u32, 4u32);
+    let y: Vec<u16> = (0..h).flat_map(|_| (0..w).map(|x| (64 + x * 876 / 255) as u16)).collect();
+    let n = (w * h) as usize;
+    let data = PixelData::Yuv16 { planes: [Arc::new(y), Arc::new(vec![480; n]), Arc::new(vec![560; n])], chroma: Chroma::C444, bits: 10, alpha: None };
+    let color = ColorInfo { matrix: Matrix::Bt2020Ncl, transfer: Transfer::Pq, primaries: Primaries::Bt2020, range: Range::Limited };
+    let frame = Arc::new(VideoFrame { width: w, height: h, data, color, par: (1, 1), pts: Default::default() });
+    let t = Arc::new(InputTransform::new(ColorSpace::from_info(&color), color.range, &ColorPipeline::REC709, None));
+    let e = effect("brightness_contrast", &[("brightness", fl(-10.0)), ("contrast", fl(20.0))]);
+    let op = FxOp::eval(&e, &cx(Tick::ZERO, 1.0), w as usize, h as usize).expect("brightness / contrast op");
+    let layer = PlanLayer {
+        fx: Some(Arc::new(LayerFx { size: (w, h), decimation: 1, ops: vec![op] })),
+        input: Some(t),
+        ..PlanLayer::new(frame, Affine::IDENTITY, 1.0, Blend::Normal)
+    };
+    let plan = FramePlan::Layers { width: w as usize, height: h as usize, layers: vec![layer] };
+    let cpu = execute_cpu(&plan).over_black_rgba8();
+    for stage in [true, false] {
+        let mut c = GpuCompositor::new(&dev, &q);
+        if !stage {
+            c.fx = None;
+        }
+        c.composite(&plan);
+        let (_, _, gpu) = c.read_output().expect("readback");
+        let (p99, mean, max) = stats8(&cpu, &gpu, &vec![true; n]);
+        eprintln!("PQ layer with effects, effect stage {stage}: p99 {p99}, mean {mean:.2}, max {max}");
+        assert!(p99 <= 6 && mean < 1.5, "stage {stage}: p99 {p99}, mean {mean}, max {max}");
+    }
+}
+
 /// The half-float upload is lossless for this input, so a 1:1 effect source draw must be too.
 /// Odd dimensions expose interpolation drift that can invent opacity before any effect runs.
 #[test]
diff --git a/crates/gpu/src/lib.rs b/crates/gpu/src/lib.rs
--- a/crates/gpu/src/lib.rs
+++ b/crates/gpu/src/lib.rs
@@ -84,6 +84,8 @@ pub struct GpuCompositor {
     accum_readback_buf: Option<(wgpu::Buffer, u64)>,
     uploads: HashMap<(usize, u32, u32), Uploaded>,
     clock: u64,
+    /// The resolve pass's display transform (`set_display_transform`), as plain numbers.
+    display: [f32; DISPLAY_FLOATS],
     dummy: wgpu::TextureView,
     /// The effect stage (None on devices without compute shaders or storage textures: layers
     /// with effects are then rendered on the CPU here).
@@ -253,6 +255,16 @@ pub fn prepare_frame(f: &VideoFrame) -> Option<Prepared> {
     }
 }
 
+/// The display transform to resolve a *sequence* plan with: `display` (the sequence's,
+/// `colorman::display_transform(..).gpu_params()`) for a plan composited in its working space,
+/// none for a [`FramePlan::Image`]. That image is `render_sequence`'s output, already converted
+/// for the monitor, and converting it again would tone map and gamut map it twice. (The CPU
+/// images the compositor makes itself, for a transition it cannot mix, are in working space and
+/// keep the transform: they are drawn inside a working-space plan.)
+pub fn sequence_display(plan: &FramePlan, display: Option<filmcraft_color::GpuOutputParams>) -> Option<filmcraft_color::GpuOutputParams> {
+    if matches!(plan, FramePlan::Image(_)) { None } else { display }
+}
+
 /// Convert every layer of a plan for upload. Thread-safe and GPU-free: run it on a worker.
 pub fn prepare(plan: &FramePlan) -> PreparedPlan {
     match plan {
@@ -314,7 +326,16 @@ impl GpuCompositor {
         });
         let final_bgl = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
             label: Some("final"),
-            entries: &[wgpu::BindGroupLayoutEntry { binding: 0, ..tex_entry(0) }],
+            entries: &[
+                wgpu::BindGroupLayoutEntry { binding: 0, ..tex_entry(0) },
+                // the resolve pass's display transform (`set_display_transform`)
+                wgpu::BindGroupLayoutEntry {
+                    binding: 1,
+                    visibility: wgpu::ShaderStages::FRAGMENT,
+                    ty: wgpu::BindingType::Buffer { ty: wgpu::BufferBindingType::Uniform, has_dynamic_offset: false, min_binding_size: None },
+                    count: None,
+                },
+            ],
         });
         let pl =
             device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor { label: Some("layer"), bind_group_layouts: &[Some(&layer_bgl)], immediate_size: 0 });
@@ -420,6 +441,7 @@ impl GpuCompositor {
             accum_readback_buf: None,
             uploads: HashMap::new(),
             clock: 0,
+            display: [0.0; DISPLAY_FLOATS],
             dummy,
             fx,
             transition: transition::PushStage::new(device),
@@ -699,11 +721,58 @@ impl GpuCompositor {
         buf
     }
 
+    /// The resolve pass's display transform uniform ([`set_display_transform`](Self::set_display_transform)).
+    fn display_buffer(&self) -> wgpu::Buffer {
+        let bytes: Vec<u8> = self.display.iter().flat_map(|v| v.to_le_bytes()).collect();
+        let buf = self.device.create_buffer(&wgpu::BufferDescriptor {
+            label: Some("display-u"),
+            size: bytes.len() as u64,
+            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
+            mapped_at_creation: false,
+        });
+        self.queue.write_buffer(&buf, 0, &bytes);
+        buf
+    }
+
     /// Composite a plan; returns the output view (sRGB, over black) and its size.
     pub fn composite(&mut self, plan: &FramePlan) -> (wgpu::TextureView, (u32, u32)) {
         self.composite_prepared(plan, None)
     }
 
+    /// Set the resolve pass's display transform (the sequence's working space → monitor,
+    /// `filmcraft_color::OutputTransform::gpu_params`). Layers are composited in the sequence's
+    /// working linear space, so an HDR / wide-gamut sequence needs it to come out right; `None`
+    /// (the default) leaves the accumulator as the monitor's own space, as for a plain sequence.
+    /// Only the resolve pass uses it: a caller that reads the working-linear accumulator back
+    /// (the export path) is unaffected.
+    pub fn set_display_transform(&mut self, p: Option<filmcraft_color::GpuOutputParams>) {
+        let Some(p) = p else {
+            self.display = [0.0; DISPLAY_FLOATS];
+            return;
+        };
+        let mx = p.matrix.unwrap_or([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]);
+        let (sp, dp) = p.tone.unwrap_or((0.0, 0.0));
+        let dl = p.compress.unwrap_or([0.0; 3]);
+        let on = |b: bool| if b { 1.0 } else { 0.0 };
+        let curve = match p.target.curve() {
+            filmcraft_color::Curve::Sdr | filmcraft_color::Curve::Srgb => 0.0,
+            filmcraft_color::Curve::Linear => 1.0,
+            filmcraft_color::Curve::Pq => 2.0,
+            filmcraft_color::Curve::Hlg => 3.0,
+            // not reachable: `gpu_params` refuses a target the shader cannot encode
+            filmcraft_color::Curve::Log(_) => 0.0,
+        };
+        #[rustfmt::skip]
+        let d = [
+            1.0, curve, on(p.tone.is_some()), on(p.matrix.is_some()),
+            on(p.compress.is_some()), sp, dp, filmcraft_color::REFERENCE_WHITE_NITS as f32,
+            mx[0][0], mx[0][1], mx[0][2], dl[0],
+            mx[1][0], mx[1][1], mx[1][2], dl[1],
+            mx[2][0], mx[2][1], mx[2][2], dl[2],
+        ];
+        self.display = d;
+    }
+
     /// [`composite`](Self::composite) with texel conversions already done by [`prepare`] (the
     /// result is identical; only the upload work on this thread differs).
     pub fn composite_prepared(&mut self, plan: &FramePlan, prep: Option<&PreparedPlan>) -> (wgpu::TextureView, (u32, u32)) {
@@ -749,7 +818,9 @@ impl GpuCompositor {
                 if self.fx.is_some() && lfx.size.0.max(lfx.size.1) <= max_side {
                     return None;
                 }
-                let img = filmcraft_render::plan::effect_image(&l.frame, lfx);
+                // decoded through the layer's input transform, as `execute_cpu` does; the image is
+                // then working-space linear, so the new layer has none
+                let img = filmcraft_render::plan::effect_image_with(&l.frame, lfx, l.input.as_deref());
                 Some(PlanLayer::new(Arc::new(VideoFrame::rgba_f32(img.w as u32, img.h as u32, img.px)), l.matrix, l.opacity, l.blend))
             })
             .collect();
@@ -858,10 +929,14 @@ impl GpuCompositor {
         if let Some(f) = self.fx.as_mut() {
             f.end_frame();
         }
+        let display_buf = self.display_buffer();
         let final_bg = self.device.create_bind_group(&wgpu::BindGroupDescriptor {
             label: Some("final"),
             layout: &self.final_bgl,
-            entries: &[wgpu::BindGroupEntry { binding: 0, resource: wgpu::BindingResource::TextureView(&accum_view) }],
+            entries: &[
+                wgpu::BindGroupEntry { binding: 0, resource: wgpu::BindingResource::TextureView(&accum_view) },
+                wgpu::BindGroupEntry { binding: 1, resource: display_buf.as_entire_binding() },
+            ],
         });
         let mut enc = self.device.create_command_encoder(&wgpu::CommandEncoderDescriptor { label: Some("filmcraft-composite") });
         // Runs of fixed-function layers draw in one pass. A layer that reads the destination
@@ -1174,6 +1249,8 @@ struct SrcInfo {
 
 /// Floats in a layer's uniform block (`U` in `composite.wgsl`: 13 vec4).
 const UNIFORM_FLOATS: usize = 52;
+/// Floats in the display transform uniform (`struct O` in `composite.wgsl`): 5 vec4.
+const DISPLAY_FLOATS: usize = 20;
 
 /// Poll a future once: wgpu's native error scopes resolve immediately. `None` if it is still
 /// pending (then the caller assumes the worst).
diff --git a/crates/render/src/colorman.rs b/crates/render/src/colorman.rs
--- a/crates/render/src/colorman.rs
+++ b/crates/render/src/colorman.rs
@@ -55,13 +55,18 @@ pub fn input_transform(cs: ColorSpace, range: Range, pipe: &ColorPipeline, peak_
 }
 
 /// The input transform of a frame the GPU compositor can colour manage itself (#965): a Y'CbCr
-/// frame of display-referred HDR or wide-gamut media (PQ, HLG, BT.2020, P3) on a plain pipeline,
-/// whose working → monitor transform is the identity. None for log media (camera-gamut
-/// compression), RGB frames and other pipelines, which [`decode`] converts on the CPU.
+/// frame whose signal → working space stages the shader runs (PQ, HLG, BT.2020, P3; log curves and
+/// camera-gamut compression are what it cannot). The sequence's working → monitor transform runs in
+/// the compositor's resolve pass (`OutputTransform::gpu_params`), so an HDR working space is fine;
+/// when the resolve pass cannot finish it, the whole frame stays on the CPU. None for log media,
+/// RGB frames and pipelines the shader cannot finish, which [`decode`] converts on the CPU.
 pub fn gpu_input(project: &Project, item: ItemId, pipe: &ColorPipeline, cs: ColorSpace, frame: &VideoFrame) -> Option<Arc<InputTransform>> {
-    if !pipe.is_plain() || !matches!(frame.data, filmcraft_frame::PixelData::Yuv8 { .. } | filmcraft_frame::PixelData::Yuv16 { .. }) {
+    if !matches!(frame.data, filmcraft_frame::PixelData::Yuv8 { .. } | filmcraft_frame::PixelData::Yuv16 { .. }) {
         return None;
     }
+    // The compositor's resolve pass finishes the frame with the sequence's display transform:
+    // when the shader cannot encode that, the whole frame stays on the CPU.
+    display_transform(pipe).gpu_params()?;
     let t = input_transform(cs, frame.color.range, pipe, source_peak_nits(project, item));
     t.gpu_params().map(|_| t)
 }
diff --git a/crates/render/src/plan.rs b/crates/render/src/plan.rs
--- a/crates/render/src/plan.rs
+++ b/crates/render/src/plan.rs
@@ -229,7 +229,13 @@ fn gpu_chain<'a>(project: &Project, item: &'a TrackItem, opts: RenderOptions) ->
 pub fn plan_frame(project: &Project, seq_id: ItemId, t: Tick, opts: RenderOptions, sources: &dyn SourceProvider) -> FramePlan {
     let Some(seq) = project.sequence(seq_id) else { return FramePlan::Image(crate::Image::new(1, 1)) };
     let (w, h) = output_size(seq, opts.scale);
-    if whole_frame_on_cpu(project, seq, t) {
+    // An HDR / wide-gamut sequence plans as layers too: its sources are decoded into the working
+    // space (the shader's input transform for display-referred media, `colorman::decode` on the
+    // CPU for log media and camera gamuts) and the compositor's resolve pass runs the sequence's
+    // display transform (working space → monitor). Frames the compositor cannot finish stay on
+    // the CPU: an adjustment layer, a transition it cannot mix, or a display transform the shader
+    // cannot encode (an HDR target).
+    if !layered_at(project, seq, t) || crate::colorman::display_transform(&seq.settings.color).gpu_params().is_none() {
         return FramePlan::Image(crate::render_sequence(project, seq_id, t, opts, sources));
     }
     let mut layers = PlanStack::default();
@@ -256,13 +262,17 @@ pub fn plan_frame(project: &Project, seq_id: ItemId, t: Tick, opts: RenderOption
 /// Whether [`plan_frame`] hands the frame at `t` back as one CPU image ([`FramePlan::Image`]). A
 /// GPU renderer gains nothing on such a frame, so a caller with only a few of them (the export's
 /// pool) can render it on the CPU without taking one.
+/// Whether the frame at `t` needs the CPU compositor: an HDR / wide-gamut sequence, an adjustment
+/// layer, or a transition the compositor cannot mix. Callers that render on the CPU only (the
+/// export's pool) use it to skip taking a GPU plan; [`plan_frame`] itself plans an HDR / wide-gamut
+/// frame as layers whenever the compositor can finish it.
 pub fn is_cpu_frame(project: &Project, seq_id: ItemId, t: Tick) -> bool {
     project.sequence(seq_id).is_none_or(|seq| whole_frame_on_cpu(project, seq, t))
 }
 
-/// The frames [`plan_frame`] does not plan: an HDR / wide-gamut sequence composites and converts
-/// on the CPU, and so does a frame with an adjustment layer or a complex transition anywhere at
-/// `t` (whole-frame fallback).
+/// The frames the CPU-only path ([`is_cpu_frame`]) covers: an HDR / wide-gamut sequence (its
+/// display transform and any log-source decode), an adjustment layer, or a complex transition
+/// anywhere at `t` (whole-frame fallback).
 fn whole_frame_on_cpu(project: &Project, seq: &Sequence, t: Tick) -> bool {
     !seq.settings.color.is_plain() || !layered_at(project, seq, t)
 }
diff --git a/crates/ui-egui/src/app/gpu.rs b/crates/ui-egui/src/app/gpu.rs
--- a/crates/ui-egui/src/app/gpu.rs
+++ b/crates/ui-egui/src/app/gpu.rs
@@ -46,6 +46,15 @@ impl FilmcraftApp {
             self.disable_gpu("device error");
             return None;
         }
+        // The resolve pass finishes a *sequence* frame: the sequence's working space → monitor
+        // (`colorman::display_transform`). A plain sequence's transform is the identity, and item
+        // frames (Source monitor, thumbnails) are already in their item's space.
+        let pipe = match key.target {
+            Target::Sequence(id) | Target::SequencePlan(id) => self.session.project.sequence(id).map(|q| q.settings.color),
+            _ => None,
+        };
+        // A CPU-rendered frame (`FramePlan::Image`) is already in monitor space.
+        let display = filmcraft_gpu::sequence_display(&plan.plan, pipe.and_then(|c| filmcraft_render::colorman::display_transform(&c).gpu_params()));
         let g = self.gpu.as_mut()?;
         if g.last_key == Some(key)
             && let Some(t) = g.texture
@@ -57,6 +66,7 @@ impl FilmcraftApp {
             return None;
         }
         let t0 = web_time::Instant::now();
+        g.compositor.set_display_transform(display);
         let ran = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
             let (view, size) = g.compositor.composite_prepared(&plan.plan, Some(&plan.prepared));
             (view.clone(), size)
diff --git a/docs/architecture.md b/docs/architecture.md
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -561,9 +561,12 @@ frame (Y'CbCr/RGB + metadata) ─► source colour space: Interpret Footage over
 - **Maths** in `filmcraft-color` (`transform`, `log`, `spaces`; formulas and sources in
   [crates/color/README.md](../crates/color/README.md)); the renderer side is `render::colorman`.
 - **Fast path.** A Rec. 709 sequence without wide gamut and media whose metadata says Rec. 709 /
-  sRGB decodes exactly as before, and its layers stay on the GPU. Log, HDR or wide-gamut media,
-  and every layer of an HDR/wide-gamut sequence, are converted on the CPU (the GPU path draws them
-  as pre-rendered images, like layers with effects).
+  sRGB decodes exactly as before, and its layers stay on the GPU. Display-referred HDR or
+  wide-gamut media (PQ, HLG, BT.2020, P3) is converted in the compositor's shader, and an HDR /
+  wide-gamut sequence composites on the GPU too: the resolve pass runs its display transform
+  (working space → monitor, `OutputTransform::gpu_params`); a display transform the shader cannot
+  encode keeps the whole frame on the CPU. Log media and camera gamuts are converted on the CPU
+  (the GPU path draws them as pre-rendered images, like layers with effects).
 - **Outputs.** `RenderOptions::working_output` returns working-space pixels (HDR exports,
   scopes); otherwise the top-level render is converted for an SDR monitor. The other colour
   effects work on display-encoded values clamped to 0..1 (they clip HDR highlights above
diff --git a/docs/testing.md b/docs/testing.md
--- a/docs/testing.md
+++ b/docs/testing.md
@@ -146,8 +146,9 @@ Each codec README has the full fixture matrix and the measured results.
   to Black at 25 %, Wipe at 50 %, Timecode / Clip Name burn-in text, graphic clips (also a built-in
   graphics template with overridden properties and per-character styles with a pinned box), and colour
   management: S-Log3/S-Gamut3.Cine footage interpreted into Rec. 709 (`log_to_rec709`) and
-  Rec. 2100 PQ footage tone mapped into Rec. 709 (`hdr_tone_map`; the footage is demo frames
-  re-encoded by the test's `Encoded` source). Each frame is compared
+  Rec. 2100 PQ footage tone mapped into Rec. 709 (`hdr_tone_map`), and Lumetri grading in a
+  Rec. 2100 PQ sequence, tone mapped for the monitor (`lumetri_hdr_pq`); the footage of both is
+  demo frames re-encoded by the test's `Encoded` source. Each frame is compared
   with `crates/golden/goldens/<scene>.png` by `filmcraft_testkit::golden`:
   **PSNR ≥ 45 dB, max abs ≤ 12, 99th-percentile per-pixel max channel difference ≤ 2** (8-bit
   sRGB levels; `Tolerance::RENDER`). On failure the actual frame and a ×8 difference image go to
@@ -176,6 +177,9 @@ Each codec README has the full fixture matrix and the measured results.
   This found a real bug: the GPU upload cache was keyed by pixel-buffer address without keeping
   the buffer alive, so a new frame allocated at a freed frame's address was drawn with the stale
   texture (fixed; regression test `upload_cache_keeps_buffers_alive`).
+- `gpu_matches_cpu_on_golden_scenes` sets each scene's display transform as the app does, so the
+  colour-managed scenes (`hdr_tone_map`, and `lumetri_hdr_pq` in a Rec. 2100 PQ sequence) are
+  converted and tone mapped on the GPU too.
 
 ## 4. UI: control channel and screenshots