Exposure And Color Grading

Keep the photographed signal scene-linear until the final image chain. One declared owner controls each exposure group, tone map, and output conversion.

1. Lock the color contract

Name the scene-linear working primaries, radiance scale, alpha convention, and every producer that enters the photographed signal. Convert irradiance through the material/lighting model before metering radiance. Apply one shared physical or perceptual radiance scale to lights, environment, atmosphere, emissive materials, bloom sources, and optical effects.

Partition targets or views into exposure-control groups. A group may share GPU state only when its radiance basis, exposure policy, and reset history are identical; an automatic group also requires the same meter mask, key, and sample schedule. Assign exactly one exposure owner, one tone-map owner, and one output-conversion owner per group.

Complete when: every photographed input has one basis and scale, and every group names its members, exposure/tone-map/output owners, and state-sharing policy.

2. Choose the cheapest meter that meets the image requirement

Choose in dependency order:

1. fixed EV for a controlled or calibrated view;
2. a stratified grid or tile sampler for ordinary global auto exposure;
3. exact full-pixel hierarchical reduction when every pixel or exact mask must

contribute;

1. a log-luminance pyramid only when another feature consumes its levels or

spatial statistics;

1. a histogram only when percentile clipping fixes a demonstrated outlier or

bimodal-lighting failure.

Tap resolved, pre-bloom HDR by default. This keeps temporal noise out of the meter and avoids bloom/exposure feedback. A different tap is an authored image policy with a regression fixture.

Read [the color-pipeline reference](references/scene-referred-color-pipeline.md#meter-implementations) when implementing sampled, exact, pyramid, or histogram metering; it contains the weighted-log equations, traffic model, and small-emitter failure tests.

Complete when: fixed EV names its value/calibration and requires zero meter source reads; otherwise one meter is selected, its source and mask are named, and each cheaper rejected option has a concrete correctness failure.

3. Build the selected exposure controller

For fixed EV, bind the authored or calibrated value directly and allocate no meter, reduction, target-publication, or adaptation state.

For automatic exposure, keep targetEV, currentEV, validity, and frame indices in GPU state. Advance adaptation every rendered frame toward the last valid target, even when the meter runs less often. Keep CPU readback diagnostic-only.

Use this producer schedule only for automatic exposure:

adapt currentEV from the last completed target
  -> render and present with currentEV
  -> reduce the new meter source
  -> publish targetEV for a later frame

Bind the source texture as a real node dependency and expose both the source frame and state frame. Initialize or clear the source before the first reduction. Read [GPU exposure state](references/scene-referred-color-pipeline.md#gpu-exposure-state) for reduction state, EV adaptation, invalid aggregates, and r185 compute semantics.

Complete when: fixed EV has zero metering/adaptation work, or an automatic frame trace proves which source produced each target, adaptation remains GPU-resident, and an invalid aggregate holds the prior valid target without a CPU substitute.

4. Handle discontinuities before presentation

Give cuts an authored hold, reseed, or fixed-EV policy. Treat a radiance basis, working-primary, quantity, nonlinear-normalization, or exposure-key change as a new exposure epoch. For a pure positive scale change Lnew = k * Lold with otherwise identical semantics, preserve the displayed product by shifting the fixed EV, or both automatic states, by -log2(k):

currentEV_new = currentEV_old - log2(k)
targetEV_new  = targetEV_old  - log2(k)

Every other incompatible change starts a new exposure epoch. Automatic exposure resets meter accumulation and reseeds adapted state; fixed exposure rebinds its authored value before the new signal is presented. Resize or DPR changes rebuild only admitted resolution-dependent meter resources and sampling coordinates. Device loss recreates and reseeds only admitted GPU state under the new resource generation.

Complete when: every cut, invalid input, basis/scale change, resize, and device-loss event maps to one conversion, hold, rebuild, or reseed action that finishes before the affected frame is admitted, or to an explicit no-op because the selected branch owns no affected state.

5. Compose one final-image chain

Use this domain order unless the LUT declares another complete contract:

scene-linear HDR
  -> exposure
  -> tone map
  -> tone-mapped-linear LUT, when admitted
  -> alpha restoration
  -> output conversion

Unpremultiply before nonlinear RGB operations and premultiply afterward; exposure preserves alpha. A scene-linear LUT needs a declared shaper. A display-encoded LUT owns the exact output primaries and transfer function and therefore sits after renderOutput().

With explicit renderOutput(), set RenderPipeline.outputColorTransform = false. Mark renderPipeline.needsUpdate = true after changing the output node or output ownership. Read [tone mapping and LUTs](references/scene-referred-color-pipeline.md#tone-mapping-and-luts) only when loading, authoring, or placing a cube.

When a tone-mapped-linear cube is admitted, read [the identity 3D-LUT example](examples/identity-3d-lut.mjs) for voxel ordering and Data3DTexture configuration. It is a correctness fixture, not a look or performance bypass.

Complete when: the graph contains one exposure multiply, one tone map, one working-to-output conversion, and—only when admitted—one LUT placement in its declared domain.

6. Prove the selected branches

Capture deterministic fixtures:

• for fixed exposure, a calibration card proving the authored EV and multiplier

with zero meter source reads or adaptation state;

• for automatic exposure, a key-gray card with the expected target EV and a

bright source entering and leaving frame with monotone target/current EV trajectories;

• for a sampled automatic meter, its mask, small-emitter, and cadence cases;
• for a histogram, its underflow, overflow, and accepted percentile interval;
• when a LUT is admitted, an identity LUT with ramps and saturated swatches in

its declared domain;

• output isolation showing exactly one tone map and one output conversion.

Measure each admitted meter or LUT as a paired graph delta after warmup on the target. GPU time is available only after post-init timestamp-query support is proven; otherwise report the timing as unavailable.

Complete when: all applicable fixtures pass, the final image is inspected, and each failed fixture identifies the meter, adaptation, LUT-domain, alpha, or output-ownership cause.

Routing

Use $threejs-image-pipeline for shared MRT, temporal history, adaptive DPR, and transient lifetime; $threejs-bloom for glare source ownership; and $threejs-visual-validation for fixed-view image evidence.