A batch ran its photos strictly one after another: each render is
row-parallel inside, but a photo's decode, its PNG/TIFF/WebP encode and
the wait on a GPU render are mostly serial, so a 32-core machine averaged
~9.5 busy cores on the CPU path and 1-2 with the GPU.
run_batch now keeps export_parallelism() photos in flight (cores / 4,
1..=4; one on wasm32) on scoped lane threads plus the calling thread.
Each holds an estimate of its working memory from a new export_gate
(twice the memory budget, separate from the decode gate it nests over),
so full-size 24 MP exports (~1.2 GB each) run two at a time by default.
Lanes run at most `lanes` photos past the last one placed, which also
bounds the finished files waiting for their turn. Paths, conflict
handling, original protection, writes, results and progress stay in
batch order (Placer), so the files are the same as one at a time; a
panic in a lane becomes that photo's error.
render_bench's batch scenario now goes through export_batch. 8 x 24 MP
JPEGs, 32 cores + RTX 4080 Laptop: CPU 3.87 s -> 2.35 s, GPU 4.21 s ->
1.82 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The earlier text quoted the 13 investigation files (held-out RMS within 0.005); the whole CC0 archive shows a worst increase of 0.013 and 7 files that need the conditional-median fallback.
The per-photo look fitted its tone curve as the camera JPEG's median per
scene-luminance bin followed by isotonic pooling. On lens-corrected, sharpened
JPEGs of busy scenes neighbouring bins disagree by more than the gap between
them, so pooling produced flat steps: 13 of 31 segments on a CC0 Z 6II sample,
flat segments in 9 of 13 fitted Nikon Z / ILCE-7 samples. Each step maps a
range of scene tones to one output tone, which renders as posterised patches
with hard edges (issue #475, also #211, #454), and the conditional medians
flatten both ends of the curve.
Quantile matching (sorted scene quantiles against sorted JPEG quantiles) is
increasing by construction and insensitive to misregistration. On the same 13
files no fit has a flat segment, all still pass the gates, held-out RMS within
0.005 of before. Regression test: a misregistered synthetic JPEG.
wgpu's default DX12 compiler (Auto) loads whichever dxcompiler.dll the DLL
search path finds first. LightCraft ships none, so it picked up another
program's copy; one without dxil.dll beside it (e.g. DXC 1.7) warns that
the DXIL is unsigned, wgpu treats that as a compile error, its
indirect-validation pipelines fail, the device is lost and the window
never opens ("Parent device is lost", exit code 1).
Every wgpu instance (window and both compute devices) now uses
lightcraft_gpu::backend::backend_options(): wgpu's env options with the
DX12 compiler pinned to FXC (d3dcompiler_47.dll, part of Windows), unless
WGPU_DX12_COMPILER asks for another.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Renders a raw's DNG ProfileGainTableMap (Apple ProRAW's local tone mapping)
when the photo's Profile option "Camera local tone mapping" is on
(`profile.camera_local_tone`, off by default: Lightroom Classic renders
ProRAW without it). Built on #427's parser; supersedes the decode-time
applier of the first version of #273.
- raw/gaintable: GainTableMap::gain / evaluator (GainEval) per DNG 1.7.1:
bilinear between tables with edges replicated, input clamp(weights .
(R,G,B,min,max))^gamma in linear ProPhoto after the baseline exposure,
index input x MapPointsN clamped to the last point. Saturating cell
math; maps whose sizes don't match their gains aren't evaluated.
SourcePlacement: where a developed source sits in the active area.
- pipeline/local_tone: the gain per output pixel in the scene-linear
stage, before the white balance, mapped back through the frame (crop,
user orientation, warp) and the EXIF orientation. lin_needs_cpu routes
the stage to the CPU when it's on (CPU and GPU match); lin_key includes
it. SourceInfo::local_tone carries the map from the decoder.
- develop: Profile::camera_local_tone (omitted from JSON when off, so
existing settings and their hashes are unchanged); Profile row checkbox
when the source has a map; translations.
- RENDER_CACHE_VERSION 17. Docs: parity row and What's New as an opt-in.
Off renders pixel-identical to main on two iPhone 15 Pro ProRAWs (12 and
48 MP). On adds ~60-75 ms to the scene-linear stage at 1536x2048 on a
2-core VM, the same for both files (cost follows output pixels).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VFqow2P1EtDxt6nRjjQoTA
An odd-height adaptive QP map (EOS R7: 4732 sensor rows, 1183 QP rows)
was rejected as unsupported, so R7 C-RAW files opened from their embedded
JPEG. The map decodes to the exact end of its entropy stream; the last
row of the second-level high-frequency bands has no partner row and its
QP is used alone. Even-height maps never reach this rule, so the six
verified M50/R100/R8 fixtures are unchanged.
Checked on paired R7 RAW and C-RAW captures (ISO 100 and 6400): all
six C-RAW files decode, developed images match their RAW counterparts
visually, and ISO 6400 sensor means agree within 0.1%.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gs6roPby7THvteijZwptQt
1.Added Simplified Chinese font support for web and desktop builds.
2.Kept Inter as the default font, with Chinese used only as a fallback.
3.Verified Chinese glyph tests and a macOS universal build.
Signed-off-by: supper_yijie <syj794844854@163.com>
Samsung SRW raws whose raw IFD (private compression 32769 or 32770) stores the samples uncompressed are decoded from the sensor data: 12-bit samples packed two in three bytes (bit order told apart from the samples; NX5, NX10, NX11 and NX20, NX200, NX210, NX1000, NX1100) and 16-bit little-endian words (EX1, WB2000). CFA layout from the raw IFD's CFAPattern; framing, white balance, black and saturation levels from the maker note (value offsets relative to the note; the vectors are stored offset by its EncryptionKey); a one-direction Exif picture size crops to the centre of the framing; black from dark rows or columns when the file states none; saturation plateau below the stated level wins. Compressed codings (32770 with fewer bytes, 32772, 32773) stay preview only with that reason. srw is imported from folders; SRW takes the file-local camera look. RENDER_CACHE_VERSION 17.
One config-dir helper (engine::config::config_dir; camera_profiles re-exports it) with both
the face-models and denoise-models defaults; hosts set up both model folders; Settings has
Faces and AI Denoise tabs; both command sets registered; both model dialogs; locale entries
from both; one [features] table per app manifest; Cargo.lock regenerated by cargo; the
DecodedSource cache cost counts both segmentation mattes and the denoised twin.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TIFF raws that describe a full-resolution CFA IFD with DNG's tags (levels, crop, AsShotNeutral) but carry no DNG version tag are read by the DNG reader (RawFormat::CfaTiff, dng::decode_as). RGGB when the file states no CFAPattern. Lossless JPEG scans with a predictor selection value above 7 are rejected with that reason and stay preview-only. .3fr and .fff are imported from folders. RENDER_CACHE_VERSION 17.
Decode every JPEG XL tile with wide (32-bit) modular buffers: the
modular_16bit_buffers header flag does not hold on real lossy files (68
of 108 tiles failed with narrow buffers).
Lossy (XYB) tiles are taken from the decoder as linear sRGB primaries
and restored to the declared primaries with a plain 3x3 matrix, instead
of the decoder's gamut-mapped, [0, 1]-clipped conversion. Lossy tiles
whose declared encoding cannot be restored faithfully (ICC profile,
non-linear transfer function, non-D65 white point, intensity target
above 255 nits, single plane) are unsupported. Lossless output is
unchanged.
A JPEG XL tile that fails to decode is now an error for the image
(the file falls back to its embedded preview) instead of a zero tile.
Render cache version 17.
The worker binary gets the never-crash deny list and forbid(unsafe_code) like
every production crate root. docs/ai-masks.md spells out who can reach the
unauthenticated loopback listener.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Resolve the docs/parity.md conflict, keep the HEIC rows. `cargo xtask ci` gets
a heif step (clippy + tests for lightcraft-codecs with --features heif), since
the workspace build leaves the feature off. The Cargo.toml comments no longer
say official builds enable it; the web one notes that wasm32 aborts on panic.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Canon maker note's AspectInfo (tag 0x009a: aspect code, cropped width,
height, left, top) holds the rectangle of the shot, measured from the
recommended image crop. Use it as the CR3 default crop, so 3:2, 4:3, 16:9 and
1:1 shots develop in the shape the photographer chose. The tag is ignored when
it is missing, does not fit, or the recommended crop itself is inconsistent.
RENDER_CACHE_VERSION 15 -> 16. The M50 C-RAW corpus sample (raw.pixls.us 2663)
was shot at 1:1 and now develops at 4000 x 4000.
Keep current upstream UI behavior and the later detector contract checks. Validate candidate models before replacing working weights and records, with a synthetic regression that failed before the fix. Remove the unrelated GPU equivalence-test relaxation. All seven quality gates pass with the pinned hosted fonts.
Generated with Codex.
Photo > Detect Faces runs the YuNet detector (a small pure-Rust interpreter over the ONNX graph, no
runtime dependency) and adds the faces it finds as unnamed face regions in one undo step. The detector
and the recognition models are opt-in downloads: Settings > Faces shows each model's licence and
training-data note, Download stays disabled until the terms are accepted, and the file is fetched over
https with lightcraft-fetch from the model's own repository at a commit-pinned address, checked against
its recorded size and SHA-256 and then installed. Detect Faces offers the detector's download the first
time it is used. No model file is part of the repository (the real-model tests read LC_YUNET_MODEL; two
ignored tests download from the real hosts).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Recover the unpublished denoise work against current upstream. Preserve later opt-in model setup, shared Rust HTTPS downloads, pending-photo resume and an Amount-independent switch. Keep original RAWs unchanged and share the current color finish between ordinary and denoised sources.
Bound public tiling inputs, refuse corrupt-cache exports, and retain synthetic/real-model coverage. Include two Windows path spelling fixes in existing tests. No model weights, face code, media or Tract runtime are included.
Validation: cargo xtask ci (all seven gates); installed RawNIND CPU/GPU agreement 1.38e-6; native CPU preview/export on a public 24 MP Sony RAW, unchanged original SHA-256.
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Generated with Codex.
Keep upstream tests and the fixed SAM worker, refresh the workspace version and parity summary, and route the model picker through the case-insensitive native filter helper. Carry the shared translated font-symbol correction; validation includes the pinned hosted fonts and all seven CI gates.
Generated with Codex.
