* fix(filters): Lens Blur reads its depth from an alpha channel, listed in Depth Map › Source as in Photoshop (#7400)
* test(filters): Lens Blur depth from an alpha channel by name or index, its errors, and the Source menu (#7400)
When the crop quad's horizon crossed the layer and most of the layer
(or its centre) lay beyond it, warp picked the source's majority side
and skipped every output pixel, so the crop came out blank. A clip
wholly on one side of the inverse horizon only ever samples that side,
so draw it; a straddling clip keeps the majority rule (#4863).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
They acted inside a circle fitted to the bounds' shorter side, so on a wide
or tall image only a central disc changed. Like Photoshop's, they now act
inside the ellipse inscribed in the bounds (selection or layer), reaching all
four edges; square bounds are unchanged.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Screen as 1 - (1 - c)(1 - a) pulls a 32-bit value above white down
towards 1.0 (2.0 with a = 0.5 became 1.5, the core became 1.0). Add
a·max(1 - c, 0) instead: the same result for c <= 1, unchanged above.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
#6462 left Mean, Variance, Standard Deviation, Skewness and Kurtosis 5–15%
slower than the old serial loop on one thread. The cause: `stat` matched on
the mode three times per pixel inside one shared loop. `combine` now picks
the statistic once and hands `combine_with` a closure with a constant mode,
so each statistic gets its own loop with the match folded away. The sample
buffers are sized to the layer count and filled by index instead of `push`,
which re-checked the capacity per sample.
Every output is bit for bit the same (121 hashes match main).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FUAqfMqVHdoeDa2y68KQQw
The fast local Laplacian filter upsampled every remapped level in full for
every reference (up to 96), but each pixel reads that level for at most two
references. `llf` now samples it only where the weight is nonzero, with
`up_at` (same terms, same order as `up`; row taps found once per row,
`mirror` only near the edges). `quantile` selects its element instead of
sorting the whole work copy (5 sorts per render, serial).
36 MP dialog, all cores: Shadows 252 -> 168 ms, Highlights 275 -> 184 ms,
all 14 sliders 627 -> 411 ms. Output hashes match main.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FUAqfMqVHdoeDa2y68KQQw
`stat` raised each deviation from the mean to a power with `f32::powi`,
with the exponent passed through the moment closure. On
x86_64-pc-windows-msvc that call goes to the C runtime's `powf`, whose
cubes and fourth powers can differ from repeated multiplication in the
last bit. On Linux the same call gives the bits of repeated squaring, as
compiler-builtins' `__powisf2` computes it. Kurtosis
came out one ULP higher on Windows in two values of the test fixture, and
`stack::tests::statistics_are_bit_identical_to_the_serial_implementation`
failed there.
The moments and Skewness's `sd³` now use `ipow`, which multiplies by
repeated squaring: `x·x²` for the cube and `x²·x²` for the fourth power.
Results on Linux do not change, and the recorded bits stay as they are.
`blocks` and `fnv` exist only for `harris_and_describe_hashes_recorded_from_the
_serial_code`, which is `#[cfg(all(target_os = "linux", target_arch =
"x86_64"))]` because the hashes it compares against go through the platform's
hypot / ln / cos / atan2 / sin_cos. Everywhere else the test is compiled out and
the two helpers are left unused, so `-D warnings` fails on `dead_code`:
error: function `blocks` is never used crates/algo/src/features.rs:704
error: function `fnv` is never used crates/algo/src/features.rs:708
They now carry the same cfg as the only thing that calls them.
CI's matrix is ubuntu-latest and windows-latest and runs clippy on both, so this
fails the Windows leg as much as a local run on macOS: `target_os = "linux"` is
false for both.
N is the whole algo-side Photomerge pipeline on 3 × 8 MP photos (register at the engine's
scale with the Auto layout, gain + vignette fit, seams, multi-band blend), 16 threads:
1042 ms before, 264 ms after, 3.9×, rounded to 4×. The engine's own glue (reading regions,
building weights, writing layers) is not in this number.
Registration and multi-band blending ran almost entirely on one thread per image (four
images on 16 threads), so most of the machine sat idle. Every output is bit for bit what
the serial code produced: a probe that hashes the public Photomerge entry points
(detect_many, match_pairs, align for every layout, photometric, seam_labels, multiband at
1 to 7 levels with 1 and 3 channels, with holes in the alpha, with images sticking out of
the canvas, plus direct harris / describe calls) matches on all 349 outputs before and
after, on 4 × 2 MP, 3 × 8 MP, 97×61, 211×157, 333×251, 320×240 two-image and 40×35 sets,
at 16, 3 and 1 threads.
- features `blur` (Harris window and BRIEF smoothing): the row pass runs rows in
parallel; the column pass cuts the image into bands of 64 columns, each holding the
`&mut` row segments of its own columns (no unsafe), run in parallel. Every column keeps
its own f32 running sum, started from its first window summed in order and updated in
the same order, so only the loop nest changes.
- `harris`: Sobel, the corner response and the non-maximum scan are parallel by rows (the
candidates are fully ordered by score, y, x afterwards, so arrival order is irrelevant).
`describe`: one descriptor per corner in parallel, kept in corner order.
- `reduce`, `expand`, `push_pull`: per-row maps in parallel. `reduce`'s column pass adds
five whole rows in tap order (same per-element sum, now vectorised); `expand` works out
its source columns once and truncates instead of calling libm `floor` (the clamped
position is never negative); `push_pull` builds each coarse row from its two fine rows in
raster order, which is the order the serial scatter-add used.
- `multiband`: grid fill per image, Laplacian difference, band accumulation (every row
walks the bands in the original order), normalisation, collapse and the coverage union
are parallel. Float additions are never reordered.
Left serial on purpose: the peak fold in `harris` (about 1 ms, an order-free max but not
worth the proof), the corner spacing grid, `seam_labels` (a dynamic-programming path whose
rows depend on the previous one), `photometric` sample gathering, and the pyramid levels
themselves (each level needs the one above).
Median of 5 (8 MP rows: 7), release, ms, interleaved before/after runs. The laptop was
busy with unrelated builds the whole time (load average 11 to 18 on 16 threads), so the
1-thread rows are noisy (a single thread is time-sliced against them) and rows can move
by 10 to 20 percent.
| Case (probe) | before (16 thr) | after (16 thr) | speedup | before (1 thr) | after (1 thr) |
|--------------------------------|----------------:|---------------:|--------:|---------------:|--------------:|
| register, 4 × 2 MP | 154 | 60 | 2.6× | 336 | 167 |
| multiband, 4 × 2 MP, 3 ch | 407 | 108 | 3.8× | 591 | 407 |
| multiband, 4 × 2 MP, 1 ch | 307 | 73 | 4.2× | 813 | 566 |
| whole pipeline, 4 × 2 MP | 437 | 160 | 2.7× | 716 | 504 |
| register, 3 × 8 MP (full size) | 327 | 124 | 2.6× | 978 | 391 |
| multiband, 3 × 8 MP, 3 ch | 2348 | 916 | 2.6× | 1742 | 1150 |
| whole pipeline, 3 × 8 MP | 1042 | 264 | 3.9× | 2264 | 1280 |
The 8 MP multiband row at 16 threads is the noisiest (minimum of 7: 959 to 224 ms, 4.3×);
the whole-pipeline row is steady. Max RSS for the 3 × 8 MP whole pipeline (the probe's
inputs are about 1.05 GB of it): 2.79 GB before, 2.79 GB after (three runs each).
Tests: the serial `blur`, `reduce`, `expand`, `push_pull` and `multiband` are kept in the
test modules as references and compared bit for bit over sizes that cross every band and
row boundary (1×1, 1-wide, 1-high, odd sizes, 63 / 64 / 65 / 129 columns, radius 0 to
wider than the image, holes, partial and zero weights, ROIs outside the canvas, 1 to 7
levels). Hard-coded hashes recorded from the old code cover multiband, push_pull,
reduce + expand everywhere (plain f32 arithmetic), and harris + describe on Linux x86-64
(they go through the platform's hypot / ln / cos / atan2 / sin_cos).
Claude-Session: https://claude.ai/code/session_01AfRHB1wHUjrJx6agBYGnVu
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Layer › Smart Objects › Stack Mode (`algo::stack::combine`) was one serial
per-pixel loop (1.0x parallelism, identical at 1 and 16 threads), and Entropy
built a HashMap per channel per pixel.
- Output pixels are computed in parallel in 4096-pixel chunks (`par_rows`, so
wasm keeps its serial path). Each chunk owns its scratch buffers.
- Each pixel uses exactly the same arithmetic in the same order as before, so
every mode except Entropy is bit-identical: 110 output hashes (10 modes x 11
cases: 3 x 24 MP, 7 layers of odd/mismatched sizes, transparency, HDR values,
1 layer, no layers) match the old code.
- Entropy: a `[u32; 256]` histogram plus a short list of the bins touched
replaces the HashMap; touched bins are summed in ascending order, so the result
is now deterministic. The old HashMap summed in random iteration order and
differed run to run (up to 1.2e-7 between two runs of the old code). New vs old
output: max abs difference 1.19e-7 (one f32 ulp near 1.0), on up to ~10% of
values in the 7-layer cases; the 3-layer 24 MP case is identical.
- Modes that never use the mean (Median, Min, Max, Range, Summation, Entropy) no
longer compute it; Median uses an unstable sort (same result under total_cmp).
3 x 24 MP, release, median of 5 interleaved before/after runs (load average
~3.5-4.5 from an unrelated build running alongside), ms:
| Mode | before | after (16 thr) | speedup | after (1 thr) |
|------------|-------:|---------------:|--------:|--------------:|
| Median | 1100 | 99 | 11.1x | 1064 |
| Mean | 308 | 36 | 8.6x | 338 |
| StdDev | 342 | 41 | 8.3x | 380 |
| Entropy | 4727 | 269 | 17.5x | 2730 |
| Kurtosis | 458 | 55 | 8.3x | 488 |
| Maximum | 368 | 36 | 10.1x | 346 |
The other modes land between 8x and 10x. Max RSS for Median is unchanged
(1.50 GB, of which 1.13 GB is the input layers).
Tests: hard-coded bit patterns recorded from the old code for every mode but
Entropy; Entropy against hand-computed values and run-to-run determinism; and a
chunk-boundary test comparing a multi-chunk image to per-pixel results.
Claude-Session: https://claude.ai/code/session_01AfRHB1wHUjrJx6agBYGnVu
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
photocraft_algo::faces::detect finds faces in an RGBA image with a score, a box and five landmarks, through a pure-Rust port of libfacedetection's YuNet network (Shiqi Yu, BSD-3-Clause; licence, NOTICE and ATTRIBUTION rows included). The weights are the library's, converted to half precision (107 KB), parsed defensively. image.detectFaces reports the document's faces to scripts and agents; it is read-only with no history step and no menu entry. The network is compiled out of the browser build, where the command reports itself unavailable. Head outputs are pinned against the C++ reference.
Part of #3603.
The local Laplacian filter behind Shadows and Highlights builds a whole
pyramid per reference value (up to 96). Three costs dominated it:
- blur_step called mirror() (an integer remainder) for every tap of every
pixel; the source rows are now found once per row and columns mirror only
within two pixels of an edge.
- up() blurred a zero-stuffed plane, three samples in four zero; it now adds
only the nonzero taps, in the same order.
- the small pyramid levels were split into one-row rayon tasks, costing more
in scheduling than the rows; planes under 32K samples run on one thread.
The output is bit for bit the same: a new test checks blur_step and up
against the plain versions on odd, even, tiny and parallel-sized planes.
Conflicts:
- distort_ui.rs: main's `drop_foreign` (#5092, #5123) closes the Content-Aware Fill workspace
too when another document becomes active or its own is closed, stopping its background
preview as Cancel does; what Apply wrote stays. New test
`the_workspace_is_dropped_when_another_document_becomes_active`.
- sr, tr, vi catalogs: both sides' rows. Dropped es's High, Low, Duplicate Layer and Contract,
which main has now with the same translations.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Photoshop's Replace Color shows the replacement colour under the Hue,
Saturation and Lightness sliders; picking one in its Color Picker sets the
sliders so the sampled colour becomes it. tone::replace_color_result gives
the swatch (the sampled colour after the shift) and tone::replace_color_shift
the slider values for a picked colour.
* Raw: measure the Nikon D90 image area (4288x2848 at (10, 10))
Aligning the decoded D90 sensor plane with the camera's own JPEG puts the
image at (10, 10) in the 4352x2868 sensor plane, giving the camera's
4288x2848 frame (10-pixel margins; the right border reads the camera's flat
255 from column 4310). Add the measured ImageArea so a D90 NEF opens on the
camera's frame instead of the uncropped plane, and make a measured profile
leave a file that carries its own colour calibration (a DNG) untouched.
* Codecs: open Cineon (.cin) film frames read-only
Cineon is the 10-bit logarithmic film-scan format DPX grew from. Read it from
the same Kodak header family: the generic header (big-endian, orientation at
192, channel count at 193, the channel-1 bits/dimensions), one or three
channels, three 10-bit samples packed into each 32-bit word on a word
boundary. The 10-bit code values open as U16 (round(c x 65535 / 1023)); the
file's log black/white points are a display choice and are not applied. A
depth other than 10 bits or a channel count other than 1/3 is Unsupported;
truncation errors rather than panicking. Read-only (no Cineon writer), so
.cin joins the open filters and Cineon read closes part of #1049's G10 gap.
* PSD: write tiff blocks without copies, inflate merged ZIP once
- ImageSourceData::to_bytes writes the global mask and global blocks
straight from the model with 4-byte alignment instead of cloning each
block (data included) to set its padding. Same bytes.
- The tiff transcoder's nested regions and blocks write into the parent's
buffer (taken and handed back, error or not) instead of building their
own output and copying it up per nesting level. Bodies that fail or are
not transcoded are truncated before being copied as stored or dropped.
Same bytes.
- PsdFile::from_bytes_and_merged decodes the merged image with the parse
when asked: a ZIP stream is inflated once (the decode's stream check is
the parse's); the parse-only check runs only when the decode fails, to
keep from_bytes' errors. PSD import uses it with merged_needed.
- PsdBuilder::composite documents that Gray/CMYK composites stay straight.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Content-Aware Fill: measure a thin hole by its thickness, not its bounds (#1366)
Removing long, thin objects (coiled guitar strings, wires) left their
outline as darker and lighter blotches and copied pieces of other strings
into the fill. The pyramid shrank until the hole's bounding box was four
patches across: for a string across the image that is a level where the
hole has swallowed its surroundings, so the matches found there copy
whatever looks like the blur, and the finer levels (refine-only for wide
bounds) keep them.
A hole at most a quarter as thick as its bounds are long now counts as
thin: the pyramid stops while it is still about two patches thick, and
that level starts from a membrane across it (over a few pixels, the
surroundings themselves) instead of the onion peel, whose exhaustive
search is what made a shallow pyramid unaffordable. Compact holes are
unchanged bit for bit.
Refs #1366
* Content-Aware Fill: every part of a thin selection keeps its surroundings (#1366)
Two more causes of ghosts when removing strings and the small knobs
selected with them:
* Auto sampling summarises the selection's surroundings by a few
clusters. Around a long selection those come from the strings'
surroundings, so a small part elsewhere (a tuning knob in a differently
coloured corner) lost its own surroundings and was filled from far away
in another colour. Around a thin selection the Auto area now always
reaches as far as the selection is thick.
* The membrane that starts a thin hole now starts a patch radius out, past
the anti-aliased fringe or halo that a selection drawn close to the
object leaves outside it.
Refs #1366
* Content-Aware Fill: peel a thin hole where that stays affordable (#1366)
Starting every thin hole from a membrane smoothed Spot Healing on film
grain (spot_healing_keeps_film_grain_through_overlapping_strokes: grain
9.0 against 12.1 around). The onion peel copies the grain; it was only
given up because its exhaustive search is too slow on the big level a
long thin hole stops at. It now starts the thin hole whenever hole cells
times sources stay within 2^28 comparisons, and the membrane only beyond.
Refs #1366
* Select Subject: keep subject parts coloured like the backdrop, leave enclosed gaps open
Users on #2788 reported Select Subject dropping clothes that are close in
colour to a plain wall and bridging the gap between an arm and the torso.
Both come from the same place: the saliency pass labels low-contrast parts
of the subject as background, the background colour model then learns
them and explains them away, and the final blind hole fill closes every
enclosed gap up to 5 % of the mask.
Two changes in crates/algo:
- A second GrabCut pass whose background model only knows the backdrop. A
colour model fitted to the image frame (minus a margin around the
salient blob, in case the subject runs out of the picture) marks every
pixel it explains as probable background; the blob and whatever the
backdrop cannot account for start as probable foreground. Only
components that touch the first pass's result survive.
- Hole filling that consults the colour models: specks (<= 0.2 % of the
mask) always fill, larger holes up to 5 % fill only when their pixels
fit the foreground model better than the background's.
Tests: two regressions (a cream-shouldered figure on a near-white wall
keeps its shoulders; a ring keeps its backdrop-coloured hole open), plus
ignored timing and picture-dump helpers. The 6 MP first run goes from
about 100 ms to about 200 ms on one core; the P29 repeat path is untouched.
* Keyboard shortcut sets: Set menu, New Set, Save, Delete, .kys Export and Summarize (K2)
Edit › Keyboard Shortcuts gets Photoshop's set management. A set is a saved copy of the
bindings (Preferences.shortcutSets, the one in use in shortcutSet; "" is PhotoCraft Defaults),
and the Set menu shows "(modified)" when the bindings differ from it. New Set…, Save and
Delete wait for OK like the rest of the dialog. Export… writes the dialog's bindings as a .kys
file (commands then tools, Photoshop's macOS key spelling), each row carrying our id as pc-id
so PhotoCraft imports its own sets back exactly, whatever the label; Summarize… writes
Photoshop's HTML list of every bound command by menu.
edit.keyboardShortcuts takes useSet (before the call's edits), saveSet, deleteSet, sets and
activeSet (after them), exportKys and summarize; every reply reports the sets, the set in use
and whether it is modified. Bad names, keys or counts are errors that change nothing.
* .kys import: tool keys and sets saved on Windows (K3)
Photoshop's <tool> rows now import: each matches our tool by name (our tool names are
Photoshop's) or by pc-id in sets PhotoCraft exported, and a letter other than the tool's own
becomes its override, while its own letter keeps the tool cycling with the others (#5036).
A set saved on Windows spells the primary modifier Ctrl where a Mac set writes Cmd and keeps
Control for the Control key. A set with no Cmd or Opt anywhere but some Ctrl is read as a
Windows set, its Ctrl as our Cmd: Ctrl+L is the Cmd+L default, not a Control-key override that
would also mean Control on a Mac. The import reports `windows`.
* cargo fmt the extension lists
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Vikram Bhaskaran <vikram.bhaskaran@outlook.com>
Co-authored-by: vangel76 <vangel1976@googlemail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Gert Pellin <gert@pellin.be>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Tanish Manku <tmanku.dev@gmail.com>
* Raw: measure the Nikon D90 image area (4288x2848 at (10, 10))
Aligning the decoded D90 sensor plane with the camera's own JPEG puts the
image at (10, 10) in the 4352x2868 sensor plane, giving the camera's
4288x2848 frame (10-pixel margins; the right border reads the camera's flat
255 from column 4310). Add the measured ImageArea so a D90 NEF opens on the
camera's frame instead of the uncropped plane, and make a measured profile
leave a file that carries its own colour calibration (a DNG) untouched.
* Codecs: open Cineon (.cin) film frames read-only
Cineon is the 10-bit logarithmic film-scan format DPX grew from. Read it from
the same Kodak header family: the generic header (big-endian, orientation at
192, channel count at 193, the channel-1 bits/dimensions), one or three
channels, three 10-bit samples packed into each 32-bit word on a word
boundary. The 10-bit code values open as U16 (round(c x 65535 / 1023)); the
file's log black/white points are a display choice and are not applied. A
depth other than 10 bits or a channel count other than 1/3 is Unsupported;
truncation errors rather than panicking. Read-only (no Cineon writer), so
.cin joins the open filters and Cineon read closes part of #1049's G10 gap.
* PSD: write tiff blocks without copies, inflate merged ZIP once
- ImageSourceData::to_bytes writes the global mask and global blocks
straight from the model with 4-byte alignment instead of cloning each
block (data included) to set its padding. Same bytes.
- The tiff transcoder's nested regions and blocks write into the parent's
buffer (taken and handed back, error or not) instead of building their
own output and copying it up per nesting level. Bodies that fail or are
not transcoded are truncated before being copied as stored or dropped.
Same bytes.
- PsdFile::from_bytes_and_merged decodes the merged image with the parse
when asked: a ZIP stream is inflated once (the decode's stream check is
the parse's); the parse-only check runs only when the decode fails, to
keep from_bytes' errors. PSD import uses it with merged_needed.
- PsdBuilder::composite documents that Gray/CMYK composites stay straight.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Content-Aware Fill: measure a thin hole by its thickness, not its bounds (#1366)
Removing long, thin objects (coiled guitar strings, wires) left their
outline as darker and lighter blotches and copied pieces of other strings
into the fill. The pyramid shrank until the hole's bounding box was four
patches across: for a string across the image that is a level where the
hole has swallowed its surroundings, so the matches found there copy
whatever looks like the blur, and the finer levels (refine-only for wide
bounds) keep them.
A hole at most a quarter as thick as its bounds are long now counts as
thin: the pyramid stops while it is still about two patches thick, and
that level starts from a membrane across it (over a few pixels, the
surroundings themselves) instead of the onion peel, whose exhaustive
search is what made a shallow pyramid unaffordable. Compact holes are
unchanged bit for bit.
Refs #1366
* Content-Aware Fill: every part of a thin selection keeps its surroundings (#1366)
Two more causes of ghosts when removing strings and the small knobs
selected with them:
* Auto sampling summarises the selection's surroundings by a few
clusters. Around a long selection those come from the strings'
surroundings, so a small part elsewhere (a tuning knob in a differently
coloured corner) lost its own surroundings and was filled from far away
in another colour. Around a thin selection the Auto area now always
reaches as far as the selection is thick.
* The membrane that starts a thin hole now starts a patch radius out, past
the anti-aliased fringe or halo that a selection drawn close to the
object leaves outside it.
Refs #1366
* Content-Aware Fill: peel a thin hole where that stays affordable (#1366)
Starting every thin hole from a membrane smoothed Spot Healing on film
grain (spot_healing_keeps_film_grain_through_overlapping_strokes: grain
9.0 against 12.1 around). The onion peel copies the grain; it was only
given up because its exhaustive search is too slow on the big level a
long thin hole stops at. It now starts the thin hole whenever hole cells
times sources stay within 2^28 comparisons, and the membrane only beyond.
Refs #1366
* Keyboard shortcut sets: Set menu, New Set, Save, Delete, .kys Export and Summarize (K2)
Edit › Keyboard Shortcuts gets Photoshop's set management. A set is a saved copy of the
bindings (Preferences.shortcutSets, the one in use in shortcutSet; "" is PhotoCraft Defaults),
and the Set menu shows "(modified)" when the bindings differ from it. New Set…, Save and
Delete wait for OK like the rest of the dialog. Export… writes the dialog's bindings as a .kys
file (commands then tools, Photoshop's macOS key spelling), each row carrying our id as pc-id
so PhotoCraft imports its own sets back exactly, whatever the label; Summarize… writes
Photoshop's HTML list of every bound command by menu.
edit.keyboardShortcuts takes useSet (before the call's edits), saveSet, deleteSet, sets and
activeSet (after them), exportKys and summarize; every reply reports the sets, the set in use
and whether it is modified. Bad names, keys or counts are errors that change nothing.
* .kys import: tool keys and sets saved on Windows (K3)
Photoshop's <tool> rows now import: each matches our tool by name (our tool names are
Photoshop's) or by pc-id in sets PhotoCraft exported, and a letter other than the tool's own
becomes its override, while its own letter keeps the tool cycling with the others (#5036).
A set saved on Windows spells the primary modifier Ctrl where a Mac set writes Cmd and keeps
Control for the Control key. A set with no Cmd or Opt anywhere but some Ctrl is read as a
Windows set, its Ctrl as our Cmd: Ctrl+L is the Cmd+L default, not a Control-key override that
would also mean Control on a Mac. The import reports `windows`.
* cargo fmt the extension lists
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Vikram Bhaskaran <vikram.bhaskaran@outlook.com>
Co-authored-by: vangel76 <vangel1976@googlemail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Tanish Manku <tmanku.dev@gmail.com>
* Warp: a thin layer stretched and bent keeps its curve on commit (#5793)
The mesh warp sized its triangle grid by the source rectangle only (one cell per 4 source px). A 3 px column scaled into a wide band and then warped got a single cell across, so the committed result was a straight, affine band while the preview showed the curve. The grid now also follows the destination length of the mapped rows and columns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Warp: module doc describes the grid sized on both sides of the map (#5793)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
warp() computed the source-centre fallback as (x0 + x1) in i32 and the
inverse-mapped tile footprint as `floor() as i32 - 3` / `ceil() as i32 + 3`.
With a strongly skewed finite affine (for example Homography
[1, 1e12, 0, 0, 1, 0, 0, 0, 1] on a 32x32 clip) the float-to-int cast
saturates at i32::MIN/MAX and the following -3/+3 overflows, panicking in
debug builds and wrapping in release. A source rect with coordinates near
1.5e9 would overflow the centre sum the same way.
Widen the centre sum to f64 and use saturating_sub/saturating_add for the
footprint, which is intersected with src_rect afterwards anyway.
Add crates/algo/tests/warp_clipped.rs covering clipped-vs-full equivalence,
huge scale allocation, the extreme-affine footprint, and the far-origin
source centre.
Refs #3446
Stretching a 1 px row or column faded out along the stretch: bicubic and
bilinear sampling read transparent pixels past the source edge, and with a
selection the edge was the whole layer's bounds (lifted pixels kept their
colour at zero alpha). Samples past the source rect now repeat the edge and
the outline is antialiased by coverage, and the lifted pixels' bounds are the
selected pixels'.
Fixes#3773
Co-authored-by: Fabio Lussi <198260595+itrycodinganditsbad@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Fibers renders from the foreground and background colours without
reading the layer, but its output area was the layer's content grown by
a zero halo: on an empty layer that is empty, so nothing was drawn (and
a following transform found nothing to transform). On a partly filled
layer only the content's bounding box was filled. It now writes the
reference bounds (the selection's, else the canvas) like the global
filters, without reading the whole layer as they do.
Reds/Yellows/Greens/Cyans/Blues/Magentas selected hue closeness times
chroma. Measured against Photoshop 2025 on 126 colours per preset
(Reds, Yellows, Blues; hue +-40 degrees, saturation 5..100 %,
brightness 33..100 %), Photoshop also multiplies by (2 - max): bright
colours were already exact, darker ones came out up to 61 levels short.
With the factor every sample matches within one level. The muted navy
#3d3f54 from the issue now gets 35/255, as in Photoshop (it was 21):
Photoshop also leaves it below 50 %.
Conflicts:
- edit_menu_cmds.rs: main's #4133 (Sample All Layers off unless the active layer has no pixels
in the window) now decides `sampleAllLayers` inside the workspace's fill plan; the job keeps
the plan-based fill, the spec keeps the workspace's parameters.
- lib.rs: both module lines (caf_ui, cam_preview).
- hu, tr, vi catalogs: both sides' rows. Dropped de's High/Low (main now has them) and hu's
older "Vanishing Point" row main reworded. Added the 15 workspace strings to sr.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Auto-Blend Stack Images: a depth map by depth from focus, for Lens Blur
`edit.autoBlendLayers` with `method: stack` and `depthMap: true` adds a hidden gray Depth layer, the layer each pixel is sharpest in (black = the bottom layer, white = the top), from `photocraft_algo::dff`: the ring difference filter as focus measure, guided-filter aggregation with the merged picture as guide, a streaming peak search with Gaussian interpolation and a confidence, edge-aware weighted least squares solved by multigrid-preconditioned conjugate gradients, and guided upsampling. Once on a layer mask it drives Lens Blur's depth map.
Closes#2067.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Depth from focus: a guide or plane of the wrong length is the empty map
Review follow-up on #2073: `depth_from_focus` is public and indexed
`guide[i]` without checking that `guide` and every `luma` plane are
`w × h` long. It now returns the empty map (depth 0, confidence 0) for a
mismatch, before anything is read, and the test covers a short guide, a
short plane and a size that matches neither.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Select Subject ran every graph cut (the GrabCut iterations and the
full-resolution band re-cut in `refine_band`, tile by tile on rayon)
through the non-cancellable `grid_cut`, so the Remove Background job only
saw Cancel/Esc before and after the whole "Finding subject" step.
Add `select_subject_with`, `grabcut_with` and `finish_region_with`, which
take an `Interrupt` and return `Err(Cancelled)`: each band tile checks it
before reading pixels and its cut runs through `grid_cut_cancellable`, so
in-flight tiles stop at the max-flow's next poll and unstarted ones never
start. Remove Background runs the step under `JobCtx::stage`. The old
functions keep their signatures and results (they pass `Interrupt::NONE`),
so `select.subject`, Object Selection and the layer menu are unchanged.
Co-authored-by: lloyd <lloyd@stackworx.io>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Edit > Puppet Warp warped only a layer's pixels and left a linked mask
where it was, so the mask no longer lined up with the content. Free
Transform, Warp and Perspective Warp (#3083) already move a linked mask.
`puppet_warp_with_mask` solves the mesh once and moves the mask on the
same mesh: the mask under the mesh is lifted, deformed like the pixels
and laid over the rest; the area it leaves takes the mask's default and
the mask away from the mesh keeps its values. With a selection the mesh
covers only the lifted pixels, so only that part of the mask moves. An
unlinked mask stays put.
Tests: the moved stripe at 8/16/32-bit, undo, an unlinked mask, a
selection, and a hide-all mask in the algo crate. The engine tests fail
before the fix.
Claude-Session: https://claude.ai/code/session_01NPyyabA5efedMqVWMu3fiq
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Replace Color: Localized Color Clusters (points, localized, range)
image.adjustments.replaceColor takes the Localized Color Clusters of
Select › Color Range: "points" (where the colour was picked),
"localized" and "range" (% of the canvas's longer side). With clusters
on, the colour only matches near the points, through the same
color_range_samples as Color Range, so the same colour elsewhere in the
photo (a near-white background next to a cream garment) is left alone.
Without "localized" nothing changes. localized without points, and
malformed points, are refused.
The dialog part follows once the canvas eyedropper of #4247 is in.
Refs #4457
* Replace Color dialog: Localized Color Clusters and Range
The dialog gets Photoshop's Localized Color Clusters checkbox and its
Range slider (greyed while the box is off), and the canvas eyedropper
(#4247) now also records where the colour was picked (points), so the
localized match centres on that spot. The command's params list
localized, range and points as real keys in Photoshop's order.
Refs #4457
select.colorRange takes {"select": "skinTones"}: a soft skin-tone model
in the YCbCr chroma plane (an ellipse along the skin run of Chai & Ngan,
1999) with a luminance gate, so near-black and near-white pixels drop
out. Fuzziness widens or narrows it as it does for Sampled Colors; the
selection modes, Invert and Sample All Layers work as for the other
presets. The dialog lists Skin Tones with Fuzziness enabled.
Detect Faces needs a face detector, which there is none of yet:
detectFaces: true is refused with a message instead of being ignored.
Fixes#4455
Measured against Photoshop 27.11 on a test chart: Blur More uses the
[1 2 1; 2 2 2; 1 2 1]/14 kernel, Find Edges the full Sobel magnitude, and
Polar Coordinates turns counter-clockwise from 12 o'clock with repeated
edges and Photoshop's sampling grid. All four now match within 2 levels.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Remove Background (and Select Subject's full-resolution refinement) could
hang at "Finding subject" on some photos. One refinement tile's max-flow
never finished, and the job ignores Cancel there.
Push/relabel's admissibility test was `dist[i] == dist[j].saturating_add(1)`.
A global relabel gives nodes reachable from neither terminal the height
u32::MAX, and MAX.saturating_add(1) == MAX, so two such neighbours counted
as admissible for each other. When f32 rounding strands excess on them
(source capacities near 3e8 have an ulp of 32), they pushed it back and
forth forever. Use checked_add so an unreachable node is never a push
target; its excess stays put and the loop ends.
The regression test is a 6-node graph that looped before and now matches
the Boykov-Kolmogorov cut.
Co-authored-by: lloyd <lloyd@stackworx.io>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
mix_selection looked up the selection one pixel at a time
(`sel.pixel(x, y)`, which allocates a Vec and walks the tile map) for
every output pixel of every tile of a filter applied through a
selection: about 24 million allocations and map lookups at 24 MP.
It now reads the selection over the tile once (`read_region`) and
indexes it. The result is bit-identical: the old per-pixel version is
the test oracle, with fully selected, unselected and partial pixels,
with and without alpha, on a tile crossing the selection's tiles.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Select > Sky: a white or grey backdrop that reaches the bottom is not sky
The colour prior counts bright, nearly neutral colour as overcast sky,
so a seamless studio backdrop starting at the top edge flooded the
whole background (79 % of a product shot). A sky never reaches the
bottom of a photo; a backdrop does. A region that also covers most of
the bottom rows and is nearly neutral (mean saturation under 0.1) is
now taken for a backdrop and nothing is selected; blue sky reaching
the bottom, as in a photo taken upward, still is.
Fixes#4634
* Sky: read the bottom rows of the mask with get()
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Free Transform and Warp sampled with Catmull-Rom (Keys a = -0.5).
Photoshop's Bicubic is the Keys cubic at a = -0.75 (Image Size already
uses it): a 200 % step edge matches within 0.5 levels (Catmull-Rom was 10
off). When a Bicubic transform reduces, the source is now resized first by
the whole local scale with Image Size's widened Bicubic, and the warp
samples it at about 1:1; the old pre-reduce only kicked in below 50 % and
by powers of two, so 50-99 % reductions aliased. Bilinear still samples
directly (it matched Photoshop exactly at 50 %).
Measured against Photoshop 27.11 Free Transform (max levels / mean):
Bicubic 200 % 10 / 0.36 -> 1 / 0.01; 50 % 48-59 / 1.0-1.2 -> 6-8 /
0.02-0.05; 150 % mean 0.59 -> 0.17. Bilinear unchanged. Bicubic
Smoother / Sharper / Automatic (not offered by Transform yet) and rotated
edges are follow-ups.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Third-party action sets that load since 3204231 now play. Four TrueGrit
Texture Supply sets (Debaser, Distress Press, Halftone Zine Machine,
Transparency) go from 223 to 3 unmapped steps of 533 and play end to end.
- .atn import, 108 -> 112 events: Select and Mask (refineSelectionEdge),
Filter Gallery Spatter (GEfc), Ripple, Crystallize; targeting the layer
mask or composite channel, loading the targeted channel, Step
Backward/Forward, removeFromSelection (it silently imported as a plain
select), Pattern Fill layers (by name), nudges in points, Layer From
Background, Grayscale/Bitmap mode objects, Image Size by resolution.
- Engine: layer.translate units pt; layer.select mode remove;
select.loadSelection channel targeted; image.imageSize keepPrintSize;
layer.new.layerFromBackground ifPresent/select; pattern.import data.
New from Clipboard keeps the copied document's resolution (was 72 ppi).
- Filters: Spatter's Smoothness smooths the displacement (torn but
attached edges, as Photoshop's); Crystallize fills a cell from its point
instead of the average, so masks keep hard edges.
- UI: .pat opens/drops into the Patterns panel (no UI import existed).
- Docs: gaps G11, file-format and target-app parity.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Feather r50 on 24MP was at ~81ms kernel / ~130ms scenario against the
100ms P27 budget. The cosine-series line() allocated a fresh
n-usize run table per row (thousands of ~48KB Vecs on a 24MP mask).
Reuse one buffer per worker thread via a thread-local; no arithmetic
touched, oracle tests (Photoshop-mapped + direct-Gaussian <=1/255)
prove identical output.
bench_24mp (24MP ellipse, best of 5, x86-64 release):
feather r50: 80.8 -> 70.7ms (-12.5%)
feather r250: 144.1 -> 133.6ms (-7.3%)
feather r1000: 301.7 -> 287.7ms (-4.6%)
Full perf_scenarios P27 with patch: p50 119.3ms (was over by ~30ms,
now over by ~19ms); remaining gap is session/snapshot overhead,
follow-up.
Co-authored-by: Brandon Thomas <echelon@gmail.com>
