`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 <[email protected]>
A windowed render dehazes with the whole frame's airlight, estimated on a 384 px proxy of the frame
(`local::frame_airlight`). Every window asked for that proxy again, so every pan of a 1:1 view and every
slider step while zoomed sampled the whole original down to it on the CPU: ~200 ms of CPU on a 64 MP raw.
The proxy (taken before white balance) is now kept for the last 4 (source fingerprint, frame, size); white
balance, the dark channel and the percentile still run on it every time. Same pixels: the GPU window of
`render_bench` is byte-identical.
`optics::fingerprint` indexed the first pixel of an empty image; it now hashes only the dimensions then.
render_bench ONLY=window GPU_ONLY=1 on a 64 MP Sony ILCE-7RM6 raw (RTX 4070 Ti SUPER, 4-CPU quota), CPU time:
pan 208-212 -> 9-11 ms, exposure change 200-204 -> 4 ms, NR change 211-217 -> 13 ms.
Co-Authored-By: Claude <[email protected]>
header_plan skipped an ARW6's strip but its first 64 KB and then said Ready, while the ARW6
header probe (Mode::Header) validates every tile: its picture header, stream tables and the
first lines of its residual stream, which lie in the skipped bytes. The probe of the partly
read buffer then failed ("ARW6 first-half row count outside 2..=7") where the whole file
succeeds: sparse_corpus failed on arw6-sony-ilce7rm6-apsc-land-c.arw, the first ARW6 file it
reads (it stops at the first failure). Like NEF, ORF and PEF, whose header probes read
samples, an ARW6 now gets HeaderPlan::Whole from the header plan. The preview plan is
unchanged: the embedded preview reads no tile, and the 13 ARW previews in the corpus are
still read from 21 % of their bytes.
Only the test reads the header plan today (the app's partial reads use preview_plan), so no
import changes. New test: a synthetic two-tile ARW6 whose second tile lies past a first one
of bulk size (Ready before, Whole now). sparse_corpus passes on the 8 ILCE-7RM6 ARW6 files.
cargo test --workspace: 3205 passed, 1 failed, visualize::tests::
all_masks_overlay_shows_every_visible_mask_at_once, which fails on main b3d392b4 too
(PR #1456 addresses it); fmt, clippy, parity, layers, assets and wasm pass.
Sources: own design.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A drag tick requested the loupe's draft and the frame ended; the draft,
done in 1.5-3 ms on the GPU, was picked up at the start of the next frame,
a frame after the slider moved. Now, in a drag, the loupe waits up to 40 %
of a frame for its draft when the last loupe render took well under that
(not with a render offload, nor zoomed in, where what changes on screen is
the window's draft). A CPU render, which would only make the frame late, is
not waited for.
JobPool::wait waits for the next finished job, which try_recv then hands
out; Renderer::poll's collection of finished jobs is split out so the wait
can turn the draft into the texture.
Co-Authored-By: Claude <[email protected]>
- drag_latency (ui-egui example): the whole app driven headless at 60 frames a second; per develop control,
the latency of each drag tick until its draft is the loupe (or, with ZOOM, the zoom window) texture, the
ticks never shown, the UI thread's frame time and the time to the full-quality render after release.
- drag_tick_bench (engine example): the UI-thread work of one drag tick outside the render.
- render_bench: ONLY=window (a 1:1 zoom window cut from the original: first window, pan, slider changes;
GPU_ONLY=1, DUMP=file for pixel A/B), ONLY=load / load1 / jpeg (+ LOADONLY=1): the steps of opening a photo.
The window rows carry no colon in their names, so `cargo xtask bench` records each of them (parser test).
- open_timeline (ui-egui example): opening raws in Detail through the real app logic, headless with the GPU.
Co-Authored-By: Claude <[email protected]>
With GPU rendering on and no graphics adapter at all (wgpu: "no GPU adapter among …", e.g. no
Vulkan loader: a `cargo run` outside `nix develop` on NixOS), every render ran on the CPU, much
slower, and only Settings ▸ Performance, Help ▸ System Info and `perf.gpuReason` said why.
The desktop app now queues one start-up notice per session once the device has been tried
(`gpu::ready()`, never blocking): "No graphics processor found: LightCraft renders on the CPU, which
is much slower. Details in Help → System Info.", on Linux with "On NixOS, start it from `nix
develop`." (→, not ▸: no installed font paints ▸.) `panels::notices::gpu_notice` decides from the
reason: nothing when the GPU was turned off (preference, LIGHTCRAFT_GPU, LIGHTCRAFT_GPU_BACKEND=off),
only a software adapter was skipped, or it failed later. Headless sessions and tests never show it
(`app.gpu_notice` is set by the desktop app only). Translated in all nine catalogs; one README
sentence in the Nix section.
Co-Authored-By: Claude <[email protected]>
MediaCache::enforce_budget never evicts the original of the photo on screen (set_on_screen), but it still
counted it. A 64 MP original (Sony A7R VI: 9984×6656 linear float, 760 MiB) nearly fills the default cache
share (768 MiB) alone, so the loop evicted everything else instead: the photo's preview and thumbnail sources,
the rendered previews and the prefetched neighbours. Every edit at 1:1 (the whole-frame render and the
thumbnail), every zoom out and every step to the next photo then decoded the raw again.
The pinned original is now left out of what is compared with the budget. The rest still fits the share; the
original sits over it, as the first commit of this PR already allowed.
Co-Authored-By: Claude <[email protected]>
Each drag tick changed the edited photo's settings, so the filmstrip asked
for a new thumbnail render every frame. For an unedited raw shown with its
embedded preview, the thumbnail-level source had never been decoded, so each
of those renders decoded the whole raw (a 64 MP ARW: over a second), several
side by side, and the loupe's drafts queued behind them.
- The edited photo's thumbnail is not re-rendered during a slider drag while
it has one; it follows once the drag ends.
- A thumbnail of a photo whose preview-level source is decoded (the one in
the loupe) is cut from that source (SourceRef::Fit) instead of decoding the
file again; not when the photo has a denoised picture.
Co-Authored-By: Claude <[email protected]>
The idle trim (3 s without renders) released the shared device copy of the original a zoom
window is cut from, so pausing between pans made the next pan upload it again: 800 MB for a
64 MP raw, ~1.4 s of "gpu sample" measured (5 ms when it is still there). While a Detail window
texture is held the renderer is not idle; once the windows are gone (zoomed out, another photo,
or another view, which now frees the Detail view's windows) the trim runs as before.
The Before window waited for the original only when the After's window job was already pending,
but side by side the Before is asked for first in the frame, so both decoded the 64 MP raw at
once; and a window asked for again while the first was still decoding (a pan, the layout
settling at startup) started another full decode beside it (three decodes seen at startup).
Any window now waits while a window job is pending and the original isn't held, and follows
from the decoded original.
A decoded 64 MP original (Sony A7R VI: 9984×6656 linear float, 797 MB) is bigger than the
default cache share (half of 1.5 GiB). MediaCache::enforce_budget evicted it as soon as the
render handed it over, so every pan of a 1:1 view and every slider move while zoomed decoded
the raw again (and lost the shared GPU copy, keyed by the CPU Arc).
The frontend now tells the cache which photo Detail shows (MediaCache::set_on_screen, set per
frame in panels::compare::prepare); that photo's original is not evicted for the budget and is
let go as soon as another photo or view is shown. Keying it on "no newer preview of another
photo" instead does not work: the neighbour prefetch stores previews after the original.
Compare and Reference View are where focus is checked at 1:1, and their panes show the same
stand-ins as the loupe: a thumbnail while the photo renders, the last render after an
edit, the magnified frame under a window not in yet. Each pane now shows the loupe's
"Rendering…" pill once that has lasted half a second with nothing new (widget
`compare:rendering:<pane>`), and its failed pill ("This photo can't be opened", or
"Original offline · no smart preview", the reason on hover) when the render failed behind
a thumbnail (`compare:renderFailed:<pane>`). Survey's small fit tiles don't.
Compare panes have no draft renders, so a slider drag on the photo would leave its pane
saying "Rendering…" for the whole drag: the photo being dragged doesn't count until the
drag ends. Each pane's wait belongs to the photo in it, so one that moves in (swap, Make
Select) starts its own. A retry of a failed render keeps the failed pill, as in the loupe.
Co-Authored-By: Claude <[email protected]>
A photo whose render failed while the loupe showed its embedded JPEG or a thumbnail kept
that stand-in up as if it were the photo; "This photo can't be opened" only showed when
nothing at all could be drawn. Now a caution pill at the "Rendering…" spot says so:
"Original offline · no smart preview" when that is why, else "This photo can't be opened",
with the reason on hover and for screen readers (widget `loupe:renderFailed`). Agents see
it as `ui.inspect` → `loupe.failed`, so one waiting for a finished loupe knows when there
won't be one.
A library change retries failed renders; while the retry of the same key runs, the pill
stays and "Rendering…" doesn't take over. The renderer remembers which photo a failure
belongs to (`failure_of`): the centre message no longer shows the previous photo's error
for a moment after a step. Tests share one render offload (`headless::RenderGate`) that
holds renders and can fail chosen slots.
Co-Authored-By: Claude <[email protected]>
Cause: a floating-point CFA DNG is read as stored (no check of its float
samples). AHD and PPG clamp each green estimate between its two neighbouring
greens with `est.clamp(a.min(b), a.max(b))`; when both greens are NaN the
bounds are NaN and `f32::clamp` panics (ppg.rs:45, ahd.rs green pass). Any
float DNG with two adjacent NaN greens crashed decode + develop. Infinities
alone did not panic but turned into NaN in the demosaic arithmetic.
Fix, at two layers:
- demosaic: a `clamp_between` helper replaces the three clamps. With finite
or infinite bounds it is the same `f32::clamp` call (bit-identical
results); with NaN bounds it returns the estimate unclamped, so the panic
is gone whatever puts NaN into the mosaic.
- DNG decode: non-finite float samples are replaced where the white level is
known. NaN and -inf read as black (0), +inf as the white level (a clipped
highlight). No sensor records NaN or infinity, so a valid file is never
touched; finite samples, negative or above white included, stay as stored.
Tests: `demosaic::tests::non_finite_samples_do_not_panic` (NaN block, +/-inf
and an all-NaN mosaic through Bilinear, PPG and AHD; plus `clamp_between`
equal to the old clamp bit for bit over finite, signed-zero and infinite
values) and `non_finite_float_samples_decode_and_develop` (a float CFA DNG,
uncompressed and Deflate, with NaN and +/-inf samples: decoded values
checked, develop with all three methods gives a finite image). Both panicked
before the fix at ppg.rs:45 / ahd.rs.
Co-Authored-By: Claude <[email protected]>
A panic inside RenderJob::run or QuickJob::run (a raw decoder or pipeline bug) unwound the
job pool's worker without sending a result. The slot stayed pending and its key was never
asked for again, so the loupe or the thumbnail rendered forever; and the pool never
replaces a dead worker, so after as many such panics as it has workers (2–6) no photo
rendered until LightCraft was restarted.
Both now catch the panic (guard::catch) and return it as the job's failure, answering the
request it came from, so the renderer records it like any failed render: no retry until
the catalog changes, and the workers live on. run() is the one place every render goes
through (the pool, export, Build Previews, contact sheets, face and AI-mask workers);
callers that had their own guard keep working, and the watermark preview's, now
redundant, is gone. The web build aborts on panic, so nothing changes there.
Co-Authored-By: Claude <[email protected]>
A raw's embedded JPEG, a thumbnail, the previous render after an edit or the magnified
frame under a 1:1 window all look like the finished photo. Once a render has been on its
way for 0.5 s with nothing new on screen, a pill at the bottom of the photo (widget
`loupe:rendering`) says so; above the Navigator when that is in the way.
The wait starts when the render starts coming (an idle app draws no frames), so a quick
render never shows it, nor does a drag whose drafts keep up; once up it stays until
nothing is coming, so slow drafts don't make it blink. It watches what is on screen: the
After or the Before, and the windows in use. Not an activity row: it says the pixels
aren't final (docs/background-tasks.md).
Co-Authored-By: Claude <[email protected]>
The 2668-entry table read off the black-box oracle's output is gone.
arw6_curve.rs now builds the 4096-entry table at first use from our own
measurement: piecewise linear through knots u_k = 128 k, c_k = u_k for
k <= 11 and 356.89 + 655.646 ln(k - 6) above, rounded half up, fitted to
the 7 same-exposure lossless (LJ92) / ARW6 pairs of the CC0 ILCE-7RM6 set.
A unit test pins the whole table (sum 27776336, weighted 92792980288),
its anchors, the identity through 1408 and strict monotonicity.
The corpus sums become regression values of the new output; the 12-bit
codes before the curve stay verified against the oracle (Phase 0).
docs/arw6-compression.md describes the curve, how it was measured, its
uncertainty, the structure and x2-scale evidence, why it is not in the
file, and the after-the-fact comparison with the oracle's curve; the
tracker and docs no longer call the output bit-exact against LibRaw.
Co-Authored-By: Claude <[email protected]>
docs/superpowers/ held the working spec and plan, which pointed at the
local plan/ folder. The ARW6 module docs now cite the sections of
docs/arw6-compression.md, which gains the facts they relied on from the
spec: the NP field of word 0, the stream header word layout and how the
first half's residual lines are counted.
Co-Authored-By: Claude <[email protected]>
`cargo xtask corpus --download` now fetches one full-frame Compressed and
one APS-C Compressed HQ file (about 94 MB) instead of all 12 (about
630 MB); the corpus test checks those two and the lossless-sibling
regression test is gone, since nothing would download those files.
docs/arw6-compression.md keeps the full table with source paths and
SHA-256, says which two xtask fetches and how to fetch the rest with
`curl -L`, and records the CC0 1.0 statement of RawDB's set record
(accessed 2026-10-10).
Co-Authored-By: Claude <[email protected]>
The RawDB entries in xtask carry their SHA-256 as the third field, which
`cargo xtask corpus --download` checks; the corpus test keeps only the
sensor sum checks.
Co-Authored-By: Claude <[email protected]>
