Files
lightcraft/docs/process-versions.md
PiMinusMax 3943279105 Colour noise reduction: joint linear chroma filter with speck brake, CPU and GPU (#815) (#1348)
* WIP: colour NR v2 prototype (env LIGHTCRAFT_NR_V2, CPU only)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: colour NR v2 on by default (LIGHTCRAFT_NR_V2=0 turns it off), GPU renders read back for it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: cache renders with the new colour NR apart from older ones

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: colour NR v2 on the GPU (fused down / apply kernels), coefficient grid sigma / 2

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: colour NR v2 on the CPU without full-size intermediates (row-streamed passes)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: report peak device-buffer use per GPU stage with LIGHTCRAFT_PROFILE

* WIP: colour NR v2 linear mode (NRV2_LINEAR=1: c - Y, additive, noise model), Jensen term, mean-drift probe

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: colour NR v2 noise model with equal-count luminance bands and a non-negative fit; probe asserts the mean over 3 noise realisations

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: colour NR v2 linear mode by default (per channel, K 64, coarse pass smooths only the coarse part); CPU on GPU renders; real-colour probe

Known: tests_window::noise_reduction_does_not_depend_on_the_window fails (the noise is measured in the
rendered window; the square-root mode failed it too).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: colour NR v2 joint guide (NRV2_JOINT=1, opt-in), luminance guide knobs, leaf-next-to-black probe, camera-look guard from 858

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: colour NR v2 reads its env knobs once (per-pixel env reads made the linear mode 11x slower); joint stage timings with LIGHTCRAFT_PROFILE

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: joint colour NR without redundant passes (fused paths + next weights, none after the last scale, delta-1 weights from the path), two-component buffers, parallel reconstruction

Bit-identical to 8b89526d (test joint_matches_the_reference_bit_for_bit against a test-only copy).
Cat 32.7 MP, CPU: 6.2 -> 5.0 s at Color 25, 7.0 -> 5.6 s at 100; working buffers 2.36 -> 1.44 GB.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: joint colour NR with separate H/V slice loops and zero-difference shortcuts (no division, weight 1 without exp)

Still bit-identical to 8b89526d. Interleaved 24 MP bench: 2.72 s vs 4.52 s for the reference.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: joint colour NR on the GPU (nrcj.wgsl), default on; whole-image noise model from samples, halo bands above 2 GB

Kernels mirror nrc_joint step for step (same scale plan, noise model fitted on the CPU from GPU samples).
Cat 32.7 MP on a Quadro M2200: wb/nr 1.2-1.3 s vs 4-5 s CPU; against the CPU max 1 LSB; bands = single.
Waits after dropping large buffers (freed memory otherwise piled up to the whole 4 GB). Cache key includes joint.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: GPU joint NR plans its bands from a budget that includes the output buffer and the device buffer limit; halo 2*sum(delta) + max delta + 8; a band that cannot fit runs only this stage on the CPU

Bands vs one piece: at most 1 LSB in a few dozen pixels (running-sum boxes round differently per band start).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: GPU joint NR budgets its bands from the free device memory where it can be asked without unsafe (nvidia-smi, amdgpu sysfs), cached 3 s; fixed 2 GB otherwise

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: docs/nr-color-v2-history.md: every approach tried, measured and dropped, speed numbers, known limits (no outlier brake yet)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: history doc: speck behaviour stated qualitatively until measured on full-size renders

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: joint colour NR speck brake: compact colour specks (chroma far past the noise, small luminance step, shift not reaching the 5x5 border) lose their colour before the guide is built; weak at Color 25, full at 100; CPU and GPU

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: joint NR CPU 5.2 -> 3.5 s (33 MP), bit for bit: guide as planes with branch-free distance rows, exponentiated paths, pilot straight into planes, speck brake in place with a min/max pre-test

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: joint NR GPU: path and weight pairs instead of quads (48 instead of 60 bytes per pixel, 33 MP in one piece), speck and pilot kernels without loop-indexed private arrays, noise fit cached for identical samples; Quadro M2200 2.15 -> 1.47 s, HD 630 5.66 -> 4.26 s

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* WIP: speck brake: compactness from the second largest border share (a line or edge crosses the 5x5 border twice; one noisy border pixel no longer protects a speck)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* #815: colour noise reduction as process V3

The joint method (linear R−Y / B−Y, edge-aware guide, speck brake) is process V3 and the latest; edits on V1 and V2 keep main's blurred chromaticity. The prototype switches (LIGHTCRAFT_NR_V2, NRV2_*) and the square-root and per-channel variants they selected are gone, with their GPU kernels; nrc_params uses the former defaults as constants, so V3 renders as the prototype did. The GPU falls back to the CPU joint filter when no band fits. vram is native only (wasm build). Docs: process-versions, nr-color-joint, parity checklist, what's new.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* #815: process routing test initialises its settings in place (clippy field_reassign_with_default)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* #815: one GPU band plan at a time, with a reserve for every render on the device

An export batch renders two photos side by side (`export_parallelism`). Each joint colour NR planned
its bands from the device memory free at that moment and counted the shared buffer pool as its own,
so two plans made at once could together ask for more than there is. Measured on a Quadro M2200
(4 GB): two 33 MP exports side by side peaked at 3761 of 4096 MiB.

The plan and its bands now run under one lock (`BANDS`), and the budget leaves `VRAM_RESERVE` for
each render on the device (`RENDERS`), not just for this one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* #815: the joint colour NR's noise fit and pilot no longer switch at a hard bound

With main's preview-scaled luminance NR (#960) the CPU and GPU renders of a 640-px preview differed
by up to 10 levels (gpu equivalence `output_spaces_match`). Their images after luminance NR differ by
9e-7 rms, but one luminance band of the noise fit sat at exactly 4 times the first fit: the CPU's
samples dropped it, the GPU's kept it, and the fitted noise came out 2.6 % apart. The pilot's axis
choice then switched at a few pixels, and the edge-aware weights carried that on.

- robust_line: a band's weight fades to 0 between 4 and 5 times the fit instead of dropping at 4
  (identical below 4).
- pilot_pick: the smoother-axis mean fades in between 0.8 and 1.25 times its bound instead of
  switching at it (identical outside), on the CPU, in nrcj.wgsl and in the test reference.

All 19 CPU/GPU comparisons of the equivalence suite are within 1 level again. The new test
noise_line_does_not_jump_when_a_band_crosses_the_outlier_bound uses the bands of that preview and
fails on the old fit (5.6 % jump).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* #815: the joint colour NR measures the whole frame's noise and is keyed by process

A window measured the noise model on the pixels it held, so a loupe window
could smooth differently from the same pixels of the full render or the
export. The model now comes from fixed tiles of the whole frame
(`lightcraft_pipeline::colour_noise`, `nrc_joint::tiles`), the same for every
window, for the CPU and the GPU, and is kept in the stage cache.

The linear stage's cache key now includes the colour filter in use, so a
process change (Update to Current Process, undo) no longer reuses the other
filter's denoised image.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* M-NR: deep windows match the whole render; the VRAM query no longer stalls renders

- A U16 / F32 window with the joint colour NR is grown by the filter's reach and
  cut back, like an 8-bit one (DeepImage::crop).
- The free device memory is refreshed in the background once the cached answer is
  stale; only the first question for an adapter waits. The joint NR asks before it
  takes the BANDS lock.
- nvidia-smi starts without a console window on Windows (CREATE_NO_WINDOW).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* M-NR: the VRAM query finds the right card and keeps one answer per adapter

- A card is found by its PCI address (wgpu device_pci_bus_id, nvidia-smi pci.bus_id,
  the amdgpu sysfs directory) where both sides report one; else by a device id that
  only one card has. Two cards of one model without an address give no answer, and
  the fixed band budget applies.
- Each adapter keeps its own cached answer, so a late background refresh cannot
  replace another adapter's answer. A refresh not done after 10 s is retried.
- The cache tests wait on signals instead of fixed sleeps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* M-NR: the VRAM cache tests no longer depend on timing

The two cache tests used a 20 ms TTL: on a slow machine a refreshed answer could go
stale again between assertions and start an extra query. They now use a 30 s TTL and
age an answer explicitly (Cache::expire, test only).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* GPU: allow joint_band's eight arguments, as color_nr_joint does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: FredApps <153173854+FredApps@users.noreply.github.com>
2026-11-06 09:44:58 +01:00

10 KiB
Raw Permalink Blame History

Process versions

Changes: revision history

Every photo's develop settings record the rendering process they are interpreted by (DevelopSettings::process, a ProcessVersion in crates/develop/src/settings.rs). Lightroom calls this the process version. It lets the look of LightCraft improve (a new default tone curve, issue #146; stronger Light sliders, issue #196) without changing a single photo someone has already edited.

Process What it changes
V1 LightCraft's rendering from before process versions existed. Frozen.
V2 Highlights and Shadows keep fine detail (issue #632): their base layer (local::tone_base) is the edge-aware filter of a separable median of log luminance (radius 0.15 % of the long edge, at most 24 px, local::BASE_DETAIL) instead of log luminance itself, so twigs, lines and texture take their surroundings' shift and keep their contrast; region edges stay sharp. CPU (local::prepare) and GPU prepare branch on local::tone_base_detail (0 = no median under V1); the base plane is cached per process (local::BaseKey). Renders without Highlights or Shadows (global or local) are the same as V1's. Tone map unchanged.
V3 Pixel-scale sharpening (issue #757): Radius uses the original photo’s pixels, Detail controls halo damping, and downsized sources retain a response-matched sharpening band. V1 and V2 keep texture-band sharpening exactly. The sharpen plane is keyed by radius and process and invalidated after noise-reduction changes. The Highlights/Shadows base and tone map are unchanged from V2.
V4 V3 plus calibrated global Contrast (per channel) and Blacks (per-channel/luminance mix) on raws with a camera tone curve. Whites behaves exactly as V2, both signs; its calibration is a planned follow-up. Other sources and local adjustments retain V2 operators. See calibrated Light operators.
V5 V4 plus: a raw's tone curve rolls off into white in stops instead of reaching it within about a stop of the sensor's white, so raising Exposure keeps the highlights the sensor recorded (issue #1632). Camera curve (the look fitted to the embedded JPEG, or a DNG's ProfileToneCurve; CameraTone): past its last knot CameraTone::apply_rolloff closes the gap to white by gap / (1 + c·(2^t − 1)) at t stops past the knot, with c matching the curve's slope there and the gap halving every stop in the end (tone::ROLLOFF_STOPS); ToneMap::camera_rolloff_hdr builds the table from it. Under V1 to V4 the curve past its last knot (fitted only to camera JPEG tones below 0.85, so 1 to 3 stops below the sensor's white on most photos) closes on white exponentially in linear light: the sensor's white renders at 8-bit 246 to 254 at Exposure 0 and 254 to 255 from Exposure +1. On the 11 CC0 corpus files with a fitted look (8 Sony, 2 Nikon, 1 Fujifilm) it now renders at 239 to 250 at Exposure 0 and 247 to 253 at +1. A DNG curve that reaches white at the sensor's white renders as before. Generic curve (raws without a camera curve, ToneMap::rolloff_hdr): from half a stop below its white point (tone::GENERIC_KNEE_STOPS) the same shoulder takes over, with the extended Reinhard curve's value and slope at the join, where V1 to V4 clip at the white point (scene 1.18 at default settings, ¼ stop above the sensor's white). The sensor's white renders at about 235, 248 and 252 at Exposure 0, +1 and +2 (V1 to V4: 239, 255, 255; the CC0 Canon R8 file and 6 CC0 DNGs without a camera curve). Content past the old clip point, such as a clipped sky, now renders up to 15 codes below white at Exposure 0 and reaches 255 about 5 stops past the sensor's white. Below the last knot or the join, JPEGs and other display-referred sources, and the camera fit itself are unchanged; CPU and GPU share the tone LUT.
V6 (the latest) V5 plus: colour noise reduction filters the linear colour differences R−Y and B−Y with the joint method (issue #815, docs/nr-color-joint.md): edge-aware weights from a denoised guide shared by both differences, anisotropic diffusion over scales, a brake for isolated colour specks; Y is kept exactly. V1 to V5 keep the blurred chromaticity (local::denoise). CPU (local::denoise) and GPU (render::denoise) branch on local::joint_colour_nr. Renders with Color 0 are the same as V2's.

The rule

A change that alters how existing settings render ships as a new process. V1 is frozen.

  • A deliberate change of look (tone model, slider response, default curve, how a stage interprets a value) goes into a new process. Photos stay on the process they have until someone runs Update to Current Process (or Reset).
  • Bug fixes that make output correct where it was broken (a crash, NaN or black output, a decoder error, CPU and GPU disagreeing) are not looks and can land in place, as before.
  • Changes to how a file is decoded or which camera colour it gets change the source, not how settings are interpreted; the process doesn't reach that stage today.

Neutral display tone map (JPEG, PNG and other display-referred sources with Contrast, Whites and Blacks at 0) is the identity up to white: the SDR shoulder no longer pulls 1.0 down to 0.9875 (8-bit 255 to 254, issue #1945). It landed in place as a bug fix, since the tone map promised identity at neutral settings; edited curves keep the shoulder and their V1 look.

AI mask edges (Object and Describe selections stored as a model's grid) follow the photo's edges when they are enlarged to the output (issue #3272): the plain enlargement spilled 1–2 px onto the background, which a raised Exposure turned into a bright outline around the subject in previews and exports. Existing AI masks become 1–2 px tighter at a subject's edges, on every process; their inside, their outside and edges without contrast in the photo are unchanged. It landed in place as a bug fix: the outline is output that isn't in the photo nor in the edit, not a look anyone chose.

Adding a process

  1. Add a variant to Process (V2), append it to Process::ALL (oldest first), give it its number in Process::version and make it Process::LATEST.
  2. The compiler then points at every match on Process. The base tone map is chosen in one place, lightcraft_pipeline::finish::base_tone (both the CPU and the GPU renderer take it from there): give V2 its own arm and leave V1's alone. A process that changes another stage branches there the same way, on s.process.process(); if that stage is cached (Plan::lin_key, the planes in local.rs), add the process to its key.
  3. Keep settings_saved_before_process_versions_load_as_v1 (crates/develop/src/tests_process.rs) passing: settings saved without the field must stay V1, not become the new latest. That is the field's #[serde(default = "ProcessVersion::legacy")]; the struct-level Default gives new settings the latest instead.
  4. Render a few procedural scenes on both processes and compare, and update docs/parity-checklist.md.

What gets which process

Process
A newly imported, merged or Lightroom-imported photo the latest
Settings saved without the field (catalogs and their snapshots, versions and history, lc:settings in XMP sidecars, browser libraries: everything from before process versions), or with a damaged value V1
Reset (develop.reset) the latest
Reset of a section or a slider unchanged
Update to Current Process (develop.updateProcess) the latest, for photos on an older one; one undo step. The menu item follows the selection; a call with ids / id acts on those photos whatever is selected
Virtual copy, duplicate the source's
Copy / paste / sync / Auto Sync / presets unchanged (see below)
crs: edits (XMP from other raw developers, Lightroom catalogs, imported presets) unchanged; a new photo has the latest
Restoring a version or history step the one it had

Copy, paste, sync and presets never carry the process: it is in no settings group, so the photos they change keep their own and interpret the pasted values with it. (Lightroom offers the process version as a separate checkbox when syncing; LightCraft's groups are coarser and keep it out.) Partial settings JSON that names process itself (develop.merge, a hand-written preset) sets it.

Lightroom edits are re-interpreted by LightCraft's mapper (crates/engine/src/crs.rs), which is written against the current rendering. crs:ProcessVersion numbers Adobe's renderer, not ours, and is ignored; a photo imported with Lightroom edits is on LightCraft's latest process, and an existing LightCraft photo that receives them keeps its own.

A number above this build's latest (a library or sidecar written by a newer LightCraft) loads, is kept when saved, and renders with the newest process this build has. Update to Current Process never moves a photo to an older process; Reset does move it to this build's latest.

A damaged value (anything but a whole number from 0 to 2^32 - 1; 2.0 counts as 2) reads as V1, like a missing field, and costs nothing else: the rest of the settings load from the catalog, a snapshot, the op log or an lc:settings sidecar as usual. The damaged value isn't kept: the photo is saved as V1.

Storage

The process is saved as "process": <number> in the settings JSON, except V1, which is left out. Settings made before process versions existed are therefore saved exactly as before, and their preview-cache keys (DevelopSettings::hash64) don't change. develop.get (MCP get_develop) always states the process.

From process V3 on, persistent thumbnail, view-render and variant keys also include the resolved rendering process. A future number can keep the same settings hash across builds while resolving to different operators, so those keys must differ. V1 and V2 omit this extra field and retain their existing cached renders. No cache-version bump is needed.

Revision history

Date Change Summary
2026-11-05 minor V5: a raw's tone curve rolls off into white in stops, so raising Exposure keeps recorded highlights (#1632)
2026-10-12 minor Calibrated Contrast and Blacks, V2 Whites, resolved-process cache keys and number coordination