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>
Docs conflicts resolved on upstream's text; MRW edits reapplied with
coverage re-measured by docs/raw/coverage.py (243 verified, 946
unverified, 221 preview). Drops stale duplicate rows upstream's docs
carried in raw-parity.md and gaps.md.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Docs recounted with docs/raw/coverage.py: 241 verified, 35 unsupported. The MCP
headless extension list now comes from engine::import::EXTENSIONS upstream, which
already carries mef.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Keep main's spots.rs implementation and preserve the catalog reader and
updated capability documentation. Readers are byte-identical to the PR.
Validation: 48 catalog and 9 spot tests pass on candidate and main.
Formatting, Clippy, parity, layers, assets and WASM checks pass.
Full CI remains blocked by the same mask-overlay test failure reproduced
on main; no assertions or tests were changed to hide it.
Refs #798.
Fitted with `lightcraft-cli calibrate` to 144 usable ILCA-99M2 ARWs (of
150, seven shooting days, Standard creative style, DRO Auto) and their
in-camera JPEGs; aggregate colour statistics only.
On 30 withheld photos (profile fitted to the other 120) mean dE76 against
the camera JPEG goes 5.11 -> 4.87; 25 of 30 improve, the worst
regression is 0.24. All 150 photos stay accepted. The public raw.pixls.us
ILCA-99M2 file, not used for fitting, stays accepted (5.57 -> 5.73).
A corpus test checks that this public sample's fit keeps exactly the
built-in profile's matrix and table; the ILCE-7CR test now shares its
body. LOOK_VERSION 5 -> 6, as the profile changes the fit of this model.
Setting up Synchronize Folder, an import review or an import made a table
of every photo in the library on the session's thread before the first
file was looked at, and naming one photo's sidecar worked out who owns
every shared sidecar name in the library. In a library of 3.5 million
photos (`scan_setup_bench`, release, a folder of 500): 5.8 s before a
Synchronize scan, 2.6 s before an import review, 1.7 s before an import,
2.4 s for each sidecar written.
- A scan keeps the library as it is (photos are shared, not copied) and
looks up the files it finds (`library_files`) and the contents it probes
(`library_hashes`) on its own thread.
- An import looks the library up with its first files
(`ImportJob::look_up_library`), from the library and the options as they
were when it was set up.
- The owners of shared sidecar names are worked out for the photos asked
about (`StemOwners::among`): all five callers name only those photos.
- Synchronize Folder names its photos' sidecars in the scan, after it has
seen that the folder is there.
Now 0.9 s, 0.05 s, 0.05 s and under 1 s. What is found, and how sidecars
are named, is unchanged: the tests hold the owners against the table of
the whole library as it was made before, and the key made without
allocating against the key, for every string of up to six of the
characters that decide it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Photos imported before focus points were read got one only through Reload from Disk, which
hashes whole files on the UI thread. Now:
- lightcraft_meta::focus::read_header reads only what the decoder looks at: IFD0 (Make,
Model), the Exif IFD, the maker note's IFD and the focus tags' values (a JPEG: its segments
up to the Exif one), walked round by round and decoded with the whole-file code from a
buffer of the file's length holding just those ranges. Bounded rounds, bytes (2 MB) and
size (1 GB). One 32 KB read per α1 / A7S III ARW; on 109 real files (83 with a point)
it gives the whole-file read's point for every one.
- A background job (crates/engine/src/focus_scan.rs) reads photos with no point from a
supported body (by the stored camera; nothing else is opened), ARW or JPEG, one at a
time in path order, with a "Reading focus points" activity row (Cancel::Flag). The UI's
per-frame focus.pump gives it the face scan's pace (paused while dragging or typing) and
starts it once per library opening. Commits every 200 photos or 3 s as system ops (no
undo step, no sidecar write), only meta.focus, only on photos that still have none at the
same path. "Looked, none" goes to focus-scanned.bin beside the library, headed by the
decoder's tag so a decoder knowing more bodies starts afresh; unreadable files are retried
at the next opening.
- photo.readFocusPoints {ids?, wait?} (Photo > Read Focus Points) runs it on demand.
Measured cold on a 5400 rpm USB drive: 200 α1 / A7S III ARWs in 1.9 s (~100 photos/s,
32 KB each), so 58,796 raws take about 10 minutes instead of re-reading 2.25 TB.
Soft-delete and remove-folder now unstack in the same undo step as Recently Deleted, so leftover members show 1/2 not 3/3. Restore-from-trash does not restack.
Co-authored-by: Cursor <cursoragent@cursor.com>
`embedded_preview_srgb` (the loupe stand-in and the grid thumbnail of an unedited raw) decoded the
largest embedded JPEG whatever the size asked for: a 512 px thumbnail of a Sony A7R VI ARW decoded
its 9984 x 6656 JPEG (5.6 MB) although the file holds a 1616 x 1080 one (0.3 MB).
`lightcraft_raw::embedded_preview_covering(bytes, min_edge)` returns the smallest of a TIFF-based
raw's JPEGs whose long edge is at least `min_edge` and that is a copy of the largest one: aspect
ratio within 1 % (not a letterboxed thumbnail of another shape) and the same number of components
(not the grey semantic masks of an iPhone ProRAW DNG); else the largest, as `embedded_preview`.
The colour-space handling stays in `decode_raw_preview` (a `covering` argument); an unsupported
raw's editing source and the camera-look fit keep the largest JPEG, so a source never depends on
its size level.
On 120 raws of every make in the corpus at 256 / 512 / 1600 px: wherever the largest JPEG is still
chosen the output is identical byte for byte; where a smaller one is (Sony, Nikon, Panasonic,
Adobe DNG) orientation matches, the size is equal or 1-2 px off, mean difference <= 4.7 / 255.
- Rebased onto #1020: its plan route (`preview_window` -> `planned_preview`) now takes the box size and picks
with `embedded_preview_covering` too, so a TIFF raw read through the sparse plan gets the whole file's JPEG.
Co-Authored-By: Claude <noreply@anthropic.com>
Read the autofocus point from Sony maker notes (FocusLocation 0x2027, FocusFrameSize
0x2037, FocusMode 0x201b, AFAreaModeSetting 0x201c, AFTracking 0x2021 and the area mode
used from the enciphered 0x9402 block), for the ILCE-1 and ILCE-7SM3, whose layout was
checked against ExifTool's output on 333 real ARWs. Manual focus, centre-only Multi
reports and unchecked bodies give no point. The box is stored on the upright photo as
Meta.focus (omitted when unknown; no catalog format change), filled at import and by
Reload from Disk.
View > Show Focus Point (Alt+F) draws it in Detail, Compare and Survey through rotation
and crop; View > Zoom to Focus Point (on by default) makes Toggle Zoom centre on it; the
Info panel shows "AF-C · Wide · Animal Eye Tracking". photo.inspect returns meta.focus.
An ILCE-7RM6 ARW with Sony's distortion table now imports with Enable
Lens Corrections on: Photo::camera_defaults already turns the correction
on for any file whose probe yields embedded lens data, and the ARW
decoder emits it since the previous commit. No new default rule is
needed, so RW2 and DNG keep theirs (on when the file carries a
correction) and stored settings are never rewritten.
Tests pin it: a synthetic ARW imports with the correction on (another
body's tables are not read: off), RW2 / DNG defaults are unchanged, and
an ARW catalogued without lens data or switched off by the user keeps
its settings across a library reopen. The Optics panel's hint drops
"(DNG lens opcodes only)" in every locale.
Co-Authored-By: Claude <noreply@anthropic.com>
New vendor::arw_lens_rm6 reads the raw IFD's lens tables (0x7037 distortion,
0x7032 vignetting, 0x7035 lateral CA, their flags) and turns the distortion
table into a one-plane WarpRectilinear in OpcodeList3, with the conventions
measured on 54 ILCE-7RM6 photos: knots at i/15 of the full-frame crop half
diagonal, S(r) = r (1 + v(r)/15000), frame-filling zoom, centred on the
default crop. Only the ILCE-7RM6 is enabled; vignetting is read, not applied.
The radial least-squares fit is split out as vendor::fit_radial_warp (RW2
results unchanged); its error gate can be limited to the radii the crop
shows, so a padded active area (lossless files) doesn't refuse a fit over
pixels that are never rendered.
It sits next to main's vendor::arw_lens (#498), which holds the ILCE-7RM4A
conventions measured against Sony's exports (unit 16384, table indexed at the
output radius); the two are selected by camera model (disjoint allowlists).
arw::decode chains this module's opcodes after arw_lens::distortion, with the
same active area and the same crop (the DNG-style default crop when the file
has one, which on the ILCE-7RM6 equals Sony's crop tags), and like it only
for CFA data (not the downsized linear-RGB lossless sizes, where the knot
spacing in full-frame pixels would not hold).
The corpus test checks the conversion on the two lossless RawDB files (CC0).
The APS-C one is pinned in xtask RAW_SAMPLES with its sha256; the full-frame
one is not (at most two pins per PR, and #625 pins two Compressed ARW 6.0
files), so its case runs only where the file is present.
RENDER_CACHE_VERSION 62 -> 63.
Co-Authored-By: Claude <noreply@anthropic.com>
Grid thumbnails looked much more denoised than the loupe. `nr_params` sized
the radii from the output's scale relative to the source buffer it was given,
not the original: a 256-px cell drawn from the 512-px thumbnail level used
half the full-size radius on pixels that each average ~23 sensor pixels, and
below 0.4x (0.35x for colour) the radius stopped shrinking at all. The
guided filter's eps (the variance treated as noise) also stayed at its
full-size value although downscaling divides the noise variance by 1/scale²,
so texture in small previews was flattened as noise.
- `SourceInfo::native_long` (filled by the engine from the catalog's photo
size, kept when the decoder supplies its own facts, carried to web
workers) and `Plan::detail_scale` = output pixels per original pixel
(from `px_per_long`, so a crop no longer changes it).
- `nr_params(s, scale)`: radii x scale with no floor, eps x scale². Below
1 px the smallest blur the box Gaussian makes is blended in by the ratio
of variances (`blur::box_sigma`); a fade under 1% skips the filter.
At scale 1 the parameters are unchanged: 1:1 views and full-size exports
render as before. The GPU takes the same parameters.
- RENDER_CACHE_VERSION 63: cached thumbnails are redrawn.
Measured on a 1536x1024 noisy synthetic scene (tests_nr_scale.rs), error
of the preview against the full-size render scaled down, as a fraction of
leaving NR off, luminance NR 50 at preview scales 0.5 / 0.33 / 0.25 /
0.167 / 0.1 / 0.05: before 0.53 / 0.70 / 1.04 / 2.26 / 3.87 / 6.07, after
0.28 / 0.37 / 0.66 / 0.90 / 0.99 / 1.00. A 256-px render from the 512-px
level now matches one from the original (mean 0.07 LSB vs 0.71 before).
photo.cycleFlag steps each selected photo's own flag unflagged -> pick ->
reject -> unflagged; photo.increaseFlag / photo.decreaseFlag move it up or
down reject < unflagged < pick without wrapping. One batch undo step, Compare
and Survey act on the active photo, Auto Advance follows, text fields keep
the key. Shown in Photo > Set Flag and Help > Keyboard Shortcuts.
