Activations stay on the device between layers (net/program.rs lowers the graph; every
operator is a WGSL kernel, one submission and read-back per run); faces of a photo go
in batches of up to 16; Settings > Faces shows each model's device and a CPU-only
switch (faces.gpu); the scan's pace limits GPU use; the GPU check is an activity row.
YuNet 58 -> 21 ms, SFace 40 -> 9.6 ms, AuraFace 295 -> 71 ms (heavily loaded machine).
Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
EOS R6 Mark II C-RAW QP maps reach 172 and 175; crx.rs accepted 131-168,
so those files opened as their preview ("Canon CRX adaptive QP value 172").
Adaptive QP now goes to 175 (quantization value 51, step 228), the highest
seen in any file; uniform subband values still stop at 44.
Checked black-box against the camera JPEGs on private files: 144 R6 II /
R7 C-RAW files decode to identical hashes where they decoded before, the 3
failing ones now decode (green follows the JPEG at r 0.98), fine detail per
QP shows no jump at 172/175, 100 % crops are clean, and 1,492 files sampled
from a 35,796-file library decode with max QP 175. No third-party decoder
source consulted. Bumps RENDER_CACHE_VERSION to 63; cr3.md and What's New.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YSGFEf87ba1zKJZb9uzddK
Wake the serving loop after protocol output fails and preserve the actual writer error. Use one process-owned stdin worker so the CLI can finish normal backend cleanup without waiting for client EOF, and terminate the output worker even when a progress hook is retained.
Closes#1527
Co-Authored-By: OpenAI Codex <noreply@openai.com>
Resolve doc conflicts in gaps, parity-checklist and raw-parity; coverage
numbers regenerated with docs/raw/coverage.py, parity summary refreshed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
docs/parity-checklist.md: main's header and summary (regenerated) with this PR's Solo Mode note and history row.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A raw imported while its format couldn't be decoded kept preview_only until a
manual Reload, so new decoders in a release never reached existing photos.
When a library opens in another app version, its preview-only photos whose
stamp differs are probed again in the background and get exactly what Reload
from Disk gives them (shared reload_ops): a file that decodes now loses the
flag, one that still doesn't gets the current reason and a fresh stamp. The
result is one journaled system batch, not an undo step; develop settings and
history are kept. Unreadable files are left for next time, photos changed
while probing are skipped, and a result for a library closed meanwhile is
dropped. A toast says how many raws now develop from their raw data.
- engine: cmd/recheck.rs (stale / probe / apply) and
library.recheckPreviewOnly {all?} -> {checked, updated, cleared}; refused
without file access, where every probe would clear every flag
- desktop: recheck::tick once per library generation, through tasks::spawn
("Checking raw files" in the activity stack)
- the preview-only notice (Edit, Info) gains a Reload from Disk button
- strings translated in all nine catalogs; docs/mcp.md and the parity row
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Recheck all 54 files in the 41-model shortlist on main 25ef65c8.
Pin 40 accepted files by SHA-256 and full decoded sensor FNV-1a,
and record their camera-image correlation. Exclude eleven candidate
models after checking identity, camera module, provenance and references.
Update generated coverage and tracking documents to 265 verified models.
Keep Nikon coding limits and the own-matrix DNG rendering issue explicit.
Co-Authored-By: Codex <noreply@openai.com>
Sony ARWs record shading, lateral chromatic aberration and distortion as a camera setting plus a radial
table in the raw image IFD (0x7031-0x7037), with copies in the encrypted SR2SubIFD (0x797c-0x7982).
They are now decoded for every body that records them and carried through RawInfo and RawImage as
`lens_metadata.sony` (types in crates/raw/src/lens_metadata.rs, decoder in
crates/raw/src/vendor/arw_lens.rs):
- Raw-IFD fields are the ones the TIFF reader parsed, which is what the image itself uses. A scan of
the directory adds diagnostics only: repeated tags, and why an entry the reader skipped was dropped
(values outside the file, an unknown type, past the reader's limits). SR2SubIFD fields are read
from their first entry. Only the exact layouts are accepted: SHORT[1] or non-negative SSHORT[1]
settings, SSHORT[17] headed by 16 for shading and distortion, SSHORT[33] headed by 32 for CA.
Nothing is allocated from a count in the file.
- The raw IFD's copy wins. The SR2SubIFD copy fills in only for a tag the raw IFD lacks, never for a
malformed one, and the two copies' agreement is recorded.
- Settings keep their number and their documented meaning, Unknown(value) otherwise (never Off).
- Framing evidence (stored frame, Sony and default crops, reference rectangle, coding, 3:2) and a
per-component status. Only Applied components become opcodes.
Applied corrections are unchanged: the ILCE-7RM4A distortion conversion receives exactly the table it
received before, and no other body or component gains a correction. The ILCE-7CR distortion misses its
camera JPEG at 24 mm, and shading and CA semantics are not validated on any body; see
docs/sony-lens-corrections.md.
Locating the SR2SubIFD, now done for every ARW that has one, selects the SR2Private offset, length and
key with the TIFF reader's own rules, which lightcraft-tiff now exposes and builds its reader on
(directory_entries, entry_span, ParseOptions::budget_for, decode_value). Only each field's first value
is decoded, so the directory's other entries cost no memory sized by the file: a 16 MiB file whose
SR2Private holds eight fields aliasing it peaks at 37 kB. DirectoryEntries keeps its layout private,
and its positions() iterator uses checked arithmetic.
Tests: synthetic decoder tests; a differential test of the SR2Private selection against the TIFF
reader's in both byte orders (0 to 8,193 filler entries, unreadable and unusual first occurrences of
each field, a used-up value budget); a dhat peak-heap test; public-corpus checks against the values
ExifTool prints, with pinned and anchored SHA-256s (LIGHTCRAFT_CORPUS_REQUIRED=1 makes absent files
fail); and an opt-in local inventory check (crates/raw/tests/sony_lens_inventory.rs).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An accelerator hook in the face interpreter takes Conv nodes; lightcraft-gpu implements it with a WGSL
implicit-GEMM convolution (any stride, padding, grouping; split-K for small maps; a direct kernel for depthwise
layers), each model's weights resident on a device of its own. After loading, detection and recognition check
the GPU in the background (agreement within 1e-3 and faster than the CPU) and otherwise stay on the CPU, which
remains the reference; the first GPU error drops it. Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Treat a zero-width feather transition as a hard edge so exact-radius pixels remain unchanged, including at zero opacity. Add actual image regressions for all spot modes, image corners and two clone output sizes.
Validation: both new regressions failed on the original implementation and pass here. Formatting, workspace Clippy, HEIF and a generated-photo app smoke test pass. The unfiltered workspace gate remains blocked by the existing all-masks overlay test tracked by #1599; later parity, layers, assets and WASM stages are pending. This is a source checkpoint, not a green release candidate.
Contributor: @eliets2
