paint.remove takes several strokes at once (`strokes`) and fills them as one job and one undo
step, so the fill sees every hole together; and `newLayer` writes the fill to a new layer above
the target, only where the hole is, leaving the target as it was (with Sample All Layers it
fills from the composite). The options bar shows Photoshop's three options; with Remove after
each stroke off, strokes collect as a tint on the canvas until Enter or the ✓ applies them and
Esc or ⊘ drops them. Strokes are dropped when the tool or document changes under them.
Session::edit reported damage only for edits that changed no pixels (same_pixels) and left
every other edit to recomposite the whole document unless the command reported its own
damage. Duplicate and paste do (#2883, #2901), but Fill, Invert, filters, Delete Layer and
the undo/redo of a duplicate, a paste or a delete all recomposited everything.
- Session::edit now reports step_damage(before, after), the same bound undo and redo use.
- step_damage bounds layers only one side has in both directions, so a removed layer counts
where it drew like an added one; the bounded_additions flag and insert_damage go, and
added_bounds becomes one_sided_bounds. A layer that is (or was) the clipping base of
clipped layers above it, an adjustment, a video layer or one excluding channels still
recomposites everything (#3090).
- The subject-mask cache is dropped only when the step can have changed pixels.
perf_scenarios on the layered 20 MP document, edits on a 1 MP layer (p50, RTX 3090, Vulkan):
delete layer 9.1 -> 1.4 ms, undo a duplicate 9.3 -> 1.6 ms, fill a 500 px square
17.4 -> 5.5 ms, Invert 16.7 -> 9.2 ms. Duplicate and paste are unchanged (already bounded).
Tests: duplicate_paste_and_delete_report_where_the_layer_draws checks that each step, its
undo and its redo report the same bounded damage and that no pixel outside it changes.
new_layers_that_change_clipping_recomposite_everything (renamed) now expects bounded damage
for a painted or empty new layer and None for a new clipping base or an adjustment; deleting
and restoring a layer reports its bounds.
With the Navigator open, every edit re-rendered the whole document into a new thumbnail on the
UI thread: a full-resolution CPU composite, 303 ms after an opacity change and 486 ms after a
brush stroke on a 20 MP layered document.
The navigator cache now records the document revision of its image. When the document is
exactly one edit ahead, with the same preview key, a bounded last_damage and the same thumbnail
size, only the damage (grown by the layers' effect reach) is rendered with
render_reduced_damage and written into the texture with set_partial; an edit with empty damage
(a rename, a selection, a new empty layer) costs nothing. Anything else renders the whole
thumbnail as before. compose gains thumbnail_size, which thumbnail_buffer now uses, so both
paths agree on the size.
On the 20 MP document: 303 -> 36 ms after an opacity change, 486 -> 171 ms after a stroke
(RTX 3090 machine; the navigator renders on the CPU).
Test: navigator_damage_updates_match_a_full_thumbnail patches a thumbnail after a translate,
an opacity change, a stroke, a blend mode change and a new layer on a document with a drop
shadow, and compares it with a full thumbnail; it fails without the effect-reach inflation.
A duplicate shares its source's tiles (Arc), but its resident texture was new, so the
compositor converted and uploaded every tile again: duplicating a 20 MP photo layer uploaded
330 tiles. When a new resident's surface shares tiles with another resident of the same role,
region, texel kind, format and default pixel, its texture now starts as a copy of that one
(Stats.tiles_shared counts the tiles taken) and the usual per-tile pass fixes up whatever
differs. Layer textures gain COPY_SRC.
The copy is recorded in the frame's encoder, after whatever brought the donor up to date this
frame. Queued writes land before the encoder runs, so both textures are marked `ordered` and
their writes go through the encoder from then on; otherwise a write to either in the same
refresh would be overwritten by the copy or leak into it. A resident whose region only moved
still keeps its own texels (keep_overlap) first.
Duplicating a 20 MP photo layer in perf_scenarios: 25.1 -> 6.3 ms, 330 tile uploads -> 0
(RTX 3090, Vulkan). Paste is unchanged: the pasted layer has new tiles.
Test: duplicates_copy_their_source_texture covers a plain duplicate (no uploads), a
duplicate and its source each painted with one tile in the same refresh, and a duplicate
missing a tile, all against the CPU composite. Without the `ordered` routing the second case
is off by 167.6/255.
The compositor submitted only between 2048x2048 page cells, so one cell of a deep stack
recorded every pass into a single command buffer: 700 full-size layers make 1403 render
passes, and wgpu-core opens two hal command buffers per pass. Metal allows 4096 command
buffers per queue, so a few thousand passes in one cell can lose the device.
run_cell now counts copies and render passes and, when a refresh may submit in between,
submits the batch after FLUSH_OPS (1000) of them, waiting for the previous batch as the
page-cell flush does. Chunk textures, uniforms and bind groups outlive the submit and the
queue keeps the order, so a chunk continues in the next command buffer. A device lost in
such a submit ends the refresh with an error instead of recording more work.
Not measurably slower on Vulkan (RTX 3090): perf_scenarios rows are within noise.
Test: many_layers_submit_inside_one_chunk (700 noise layers on a 64x48 document) needs
flushes > 0 inside one chunk and matches the CPU composite; main records all 1403 passes
without a flush.
Flatten All Layer Effects baked the layer (including its masks) into the pixels and then cleared mask and vector_mask, destroying unrelated mask pixels and their enabled/disabled state permanently after save. Bake the effect visuals with the masks left off the render and keep the original mask and vector_mask objects, so each mask still applies once to the baked raster and the command stays a visual no-op.
Tests: cargo test -p photocraft-engine --lib file_cmds; cargo test -p photocraft-engine --lib; cargo clippy -p photocraft-engine --all-targets (with the two pre-existing lints allowed); cargo test -p photocraft-engine --test panic_hunt -- --ignored
PhotoCraft treats 32-bit documents as linear light (Transfer::for_document,
Exposure, Photo Filter, PSD import), but Image > Mode > 32 Bits/Channel only
rescaled the samples, so a converted 8-bit grey 128 stayed 0.502. Photoshop
stores 0.2159 (measured in Photoshop 27.11: 32-bit float TIFF of an sRGB
chart, every sample = the sRGB-to-linear value). The conversion now decodes
colour samples through the document's tone curve after widening to 32 bits,
and encodes them before narrowing (so no linear value is quantised to 8 or
16 bits). Alpha, masks and alpha channels keep their values. RGB and
Grayscale only (the 32-bit modes).
With it, Photoshop's 32-bit results match exactly for the conversion itself,
Levels, Black & White (default), Brightness/Contrast and Colorize (were 40-80
levels off in sRGB terms).
Adds Surface::map_color (colour samples only, including the default pixel)
and makes Transfer::decode/encode public.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Photoshop's absolute mode scales each range's ink change by how much the
ink can still move: adding ink by min(1, 2 x (1 - ink)), removing it by
min(1, 2 x ink). PhotoCraft added the full percentage and clipped, so
Neutrals Cyan +50 darkened a grey 64 to 0 instead of 32. Measured in
Photoshop 27.11 over every range x ink at +-50 (72 cases): all within 1
level now (8 of 72 within 3 before). CPU and GPU paths.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* engine: expose workspace.enableNarrowOptionsBar through a Workspace helper
* ui: the options bar honors the narrow options bar preference
* docs: scorecard, settings that do nothing 30 -> 29 of 154
Photoshop CS2 writes `phfl` version 3 with "4 bytes each for XYZ color".
The parser read those values as Lab x100, so real files (L = 241.91)
fell back to Unsupported. Read them as D50 XYZ with 1.0 = 32768 (the
ICC 16-bit PCSXYZ encoding; the scale is inferred from a real file, the
spec does not give it). Writing stays version 2 RGB.
Fixes#3777
Co-authored-by: Fabio Lussi <198260595+itrycodinganditsbad@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* text: missing glyph protection can turn the script fallback stack off (#1909)
* engine: apply type.missingGlyphProtection to the live text engine
* docs: scorecard, settings that do nothing 30 -> 29 of 154
Flatten Image, Merge Visible and Merge Down now composite a CMYK stack of
plain Normal raster layers in its ink values, so a lone Background flattens
bit-exact and 16-bit depth is kept. Stacks that still need the RGB
compositor separate back through the document's own CMYK profile instead of
the built-in one, and Merge Down's scratch document keeps the profile.
Fixes#3778
Co-authored-by: Fabio Lussi <198260595+itrycodinganditsbad@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Photoshop's Brushes panel nests folders by dragging one onto another; ours
could only reorder a folder among its siblings.
- brush.presets.moveGroup takes `into: {group?, folder?}`: the group or
nested folder (with everything in it, presets in order) moves inside that
folder of that group (`folder` [] = directly in the group), or with no
`group` out to the top level as a group of its own. `before` names a
sibling folder (or top-level group) there to land before. Moving a folder
into itself or one of its own folders, a name clash at the destination,
a result deeper than MAX_FOLDER_DEPTH, the ungrouped presets, `index`
with `into`, and bad shapes are errors that leave the library unchanged.
The preset store follows the move (a group moved into another one loses
its group file; a folder moved out gets one).
- Brushes panel and Brush Preset picker: a group or folder header dropped
on another header goes inside it; on the header's upper third it lands
before that folder, reordering siblings or moving a nested folder next to
it (out to the top level before a group). Every drop is one moveGroup
command (`folder_drop` maps the drop).
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A `history.*` namespace exists (history.deleteState, history.newDocument) but not history.undo, so a client or agent guessing `history.undo` gets a hard `unknown command`. Register history.undo and history.redo as aliases of edit.undo and edit.redo, with no menu entry and no shortcut; the Edit menu items keep Cmd+Z / Cmd+Shift+Z.
Tests: cargo test -p photocraft-engine --lib history_cmds; cargo test -p photocraft-engine --test panic_hunt -- --ignored; cargo xtask parity.
* fix(adjust): Hue/Saturation saturation, lightness order and Colorize as in Photoshop
Measured in Photoshop 27.11:
- Raising Saturation divides it by (1 - s): +50 doubles it. PhotoCraft
multiplied by (1 + s), up to 26 levels off.
- Master Lightness applies first; hue and saturation then work on the
result. Fixes the corpus file adjustments/rgb8/hue-saturation (Hue
+40, Saturation +25, Lightness -10): 17 levels / 11.8 % bad -> 2 / 0 %.
- Colorize builds the colour at the lightened lightness, so black with
Lightness +30 is a 30 % lightness colour, not grey (was 78 off).
- A colour range's positive Lightness lightens towards the pixel's
brightest channel, not towards white (Blues +50 was 71 off).
All 18 Hue/Saturation probe cases are within 4 levels. Corpus: the
Colorize files go to 1-3 levels (psd-tools huesaturation_colorize_rgb
120 -> 3). CPU and GPU paths.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(adjust): a range's negative Lightness darkens towards the darkest channel
Measured in Photoshop 27.11 (Blues -50 / -100, Reds -50): (102, 128, 204)
at Blues -50 becomes (102, 115, 154), towards the pixel's darkest channel,
not towards black. Mirrors the positive rule (towards the brightest).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Photoshop's Black & White tint keeps the grey's luminosity and takes the
tint's hue and saturation (SetLum + ClipColor, luma 0.30/0.59/0.11), the
same as the Color blend mode. PhotoCraft mixed half the grey with twice
the grey times the tint, up to 141 levels off. Measured in Photoshop
27.11 on 9 cases (4 tint colours over a grey ramp and hue sweeps): all
within 3 levels now. Corpus: photoshop adjustments/rgb8 and rgb16
black-white-tint go from 69 levels max / 77 % bad pixels to 1-3 / <=0.5 %.
CPU and GPU paths.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Image › Adjustments › Shadows/Highlights follows Photoshop 25.4, measured as a
black box: per-channel screen curves (1 - (1 - v)^q blended over q = 1, 2, 4,
8), the Tone mask as a Beta distribution function of an 8-bit neighbourhood
brightness, an edge-aware (guided bilateral) neighbourhood, the luma of the
result with the colour differences scaled by Color, and the Midtone S-curve.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The per-file colour fit (`fit_preview_look`) converted the embedded JPEG to sRGB only when the JPEG
carried an ICC profile, and otherwise read it as sRGB. A Sony body set to Adobe RGB writes no profile
into its previews; it states the colour space only in the raw's EXIF (ColorSpace Uncalibrated with
InteropIndex "R03", the DCF option file for Adobe RGB). The fit then targeted a desaturated,
hue-shifted reference, and the developed raw came out that way.
`photocraft_raw::embedded_preview_color_space` reads that EXIF (sRGB, Adobe RGB or unstated; never for a
DNG, whose previews are the converter's), and `preview_profile` uses it when the preview has no ICC
profile of its own. A preview's own profile still wins. LightCraft reads the same tags for the same
reason (storytold/lightcraft#778).
Measured on 19 private Sony ILCE-1M2 ARWs, all shot in Adobe RGB, against the camera's JPEG read as
Adobe RGB (median CIE76 ΔE at 240 px): full-frame 400 mm shots against blue sky 5.6-7.0 → 1.5-2.4,
their mean a* offset +4.3…+4.9 → +0.7…+1.1; every file improved, mean of the medians 4.9 → 2.2.
Tests: `exif_colour_space_of_the_previews` (sRGB, Adobe RGB; R98, a missing interop IFD, other values,
a DNG and junk are unstated) and `a_preview_without_a_profile_takes_the_raws_exif_colour_space` (fails
without the EXIF fallback; the preview's own ICC profile wins). The hostile-input harness calls the new
function.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Imported brushes whose Texture uses Height or Linear Height at a low depth
(common in Photoshop brush packs) painted nothing: the modes were modelled
backwards (paint only on the brightest texels, more of it with more depth).
Measured against Photoshop with single dabs over a 0..255 ramp pattern:
- Height: paint reaches 12 x depth x coverage into the pattern and fills
what lies below it (dark valleys first; a brighter pattern takes paint
away, which is why the reporter had to flip Brightness's sign).
Linear Height grades the shallow end by the pattern instead.
- Subtract takes depth x pattern away (white removes the most paint);
Color Dodge/Burn scale the pattern toward their neutral with depth.
These now live in texture_combine; the Dual Brush keeps mask_combine.
- Brightness shifts the pattern by value/255 (not value/150); positive
Contrast steepens by 1/(1-c/100) after the shift, negative flattens by
1+c/100 before it. Texture::level is shared by the renderer and the
Brush Settings swatch.
- ABR import mapped textureContrast with /50 for both signs; positive
values are now /100, matching the model and the Brush Settings field.
Built-in Subtract presets get Invert so they look as before.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Photoshop keeps Stylistic Sets, Contextual Alternates, Swash, Fractions, Ordinals, Titling and
Stylistic Alternates and Ornaments only in a PSD's Txt2, not in EngineData. Import now reads them
into the runs' font features, so the layout drawn while the type is edited keeps Photoshop's
glyphs instead of falling back to the default ones.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Photoshop 27.11 draws Satin Distance 0 as a 1 px offset (an edge band of
2/size), not as two copies that cancel out. PhotoCraft drew nothing. Measured
on a #808080 square with a Linear contour at sizes 5, 10 and 20; the other 12
cases already matched within 1 level. Applied to the CPU and GPU paths and
to the effect reach.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(automation): warn when text falls back from a missing font (#3391)
* cargo fmt
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>
* View: remember Extras, Show and Snap To across restarts (#3762)
View › Extras, the View › Show items (Pixel Grid, Grid, Guides, …) and
the View › Snap To targets reset to their defaults on every launch. They
are now saved in Preferences (`viewExtras`) as they change and restored at
startup, as Photoshop keeps them between sessions.
Only items that differ from their defaults are stored, so a default
changed in a later release still reaches users who never touched it. A
hand-edited file can't break them: unknown keys and non-boolean values
keep their defaults.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* View: rustfmt the view-extras restart test (#3762)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(engine): Ctrl+, toggles a layer's visibility and / its transparency lock
Photoshop's Ctrl+, (Cmd+, on macOS) hides the selected layers and, once
they are all hidden, shows them again: its Layer menu item then reads
Show Layers. layer.hideLayers now does the same, so the key toggles: it
shows them only when every targeted layer is hidden.
The / key toggles Lock Transparent Pixels, as the Layers panel's first
lock button does (layer.toggleLockTransparency). With several layers
selected, any unlocked one means lock them all; the Background, which
can't take this lock, is left out. Each press is one undoable step.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(engine): Ctrl+\ targets the layer mask and Ctrl+Alt+2 loads luminosity
Photoshop's Ctrl+\ (Cmd+\) targets the active layer's mask, as clicking
its thumbnail does, beside Ctrl+2 for the composite
(channel.target.layerMask; greyed when the layer has no mask). Its
Ctrl+Alt+2 (Cmd+Option+2) loads the image's luminosity as a selection,
the Ctrl-click on the composite channel (select.loadLuminosity).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(engine): Ctrl+Alt+Backspace fills from history
Photoshop's Ctrl+Alt+Backspace (Cmd+Option+Delete) fills the selection, or
the layer, from the History Brush's source state. edit.fillHistory is
edit.fill with contents "history": the oldest state still held, unless
"state" names another. It joins the other fill keys, so it honours the
Channels panel target and appears with them in Keyboard Shortcuts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(ui): Tab renames the next layer; Alt-click a thumbnail fits the layer
While a layer name is being edited, Tab commits it and starts renaming the
layer below, and Shift+Tab the layer above, as in Photoshop; the
Background, which isn't renamed in place, ends it. Alt-clicking a layer's
thumbnail selects the layer and fits its content on screen (View › Fit
Layer(s) on Screen).
Key tests cover Ctrl+,, /, Ctrl+\, Ctrl+Alt+2 and Ctrl+Alt+Backspace
through the shortcut dispatcher. docs/ui-parity.md lists the new keys.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ui): an explicit channel target drives the Layers panel's mask target
After Ctrl+\ targeted the layer mask, Ctrl+2 moved the Channels target back
to the composite, but the Layers panel kept the mask targeted (its own
flag, which only ever followed the engine towards the mask), so painting
still went to the mask. The same happened after clicking the mask
thumbnail. Any channel.target command (Ctrl+2, Ctrl+\, Ctrl+3…) now sets
the panel's mask and vector-mask targets from the engine's target.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Use Document::is_background after #4036 removed extra_cmds::is_background
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
engine.commands, command.list and the MCP command_list tool now give each
row a disabledReason (null when enabled, else the failed precondition) and
needsDocument (true when the command is off because no document is open),
so an agent can tell "open a document first" from "not runnable". enabled
keeps its meaning. Both listings share one row builder in the engine.
Fixes#3664
Co-authored-by: Fabio Lussi <198260595+itrycodinganditsbad@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
U8/U16 depth changes and pure channel-copy layout changes (alpha add/drop, gray to RGB) skip the two intermediate f32 buffers. Output is bit-identical to the f32 route.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A second `auth` frame on an authenticated TCP connection fell through to
the method dispatcher and came back as an unknown method, in both the
headless serve server and the desktop control server. A shared
`security::repeat_auth_reply` now answers it with `already authenticated`
(keeping the request id) and the connection keeps serving.
Fixes#3668
Co-authored-by: Fabio Lussi <198260595+itrycodinganditsbad@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
When drawing faux-bold glyphs, the offset dilation was composed as
`Xform(ox * r, oy * r).mul(&glyph)`. Because `Xform::mul(o)` applies `o`
first (`self ∘ o`), the translation offset was added after the glyph's
scaling transform, leaving faux bold dilation unscaled by horizontal
and vertical font scaling.
Applying the glyph transform to the offset transform (`glyph.mul(&offset)`)
ensures dilation is performed in font-local coordinates and scaled
proportionally with anisotropic glyph scaling, matching Photoshop.
A tiny file declaring a huge image reserved the full output buffer before any
data was checked. Add Limits::check_plausible (compression-ratio guard) and a
fallible alloc_zeroed, use them in PNG and WebP, and lower the 32-bit default
max_alloc to 1 GiB.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Photoshop's Background is the bottom layer of a file without transparency: a pixel layer that is
transparency- and position-locked, whatever the file's language calls it ("Hintergrund", "背景",
"Arrière-plan"). We matched the name "Background" in compose (deep / shallow knockout), the
engine (cut/clear, merge visible, transforms, warps, distort, erasers, styles, masks, crops,
straighten, smart-object unpack) and the UI (Properties, layer row, tab title), so a localized
file's deep knockout revealed transparency instead of the Background.
- `Document::background` / `background_mut` / `is_background` and `Layer::has_background_traits`
in the doc model replace every name check; `extra_cmds::background_to_layer` is the one
"Background → Layer 0" conversion.
- PSD import marks the bottom record position-locked when it is transparency-protected and has
no transparency channel, with or without an `lspf` block. Affinity import and HDR Toning now
lock the Background's position like every other place that creates one.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
- DisplayInfo kinds 0/1 (selected / masked areas) now import as alpha channels instead of black
spot inks; a Multichannel composite still prints them as black ink, so the look is unchanged.
- Saving a Multichannel PSD writes global light, grid and guides, saved / work / clipping paths,
EXIF and every other imported resource (print settings, IPTC, layer-id seed, alpha ids while
they match the channels), except the ones that only describe layers, plus the document-level
tagged blocks (4-byte padded). Path writing and global-block padding are shared with the
layered writer.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(io): open Krita .kra composites as editable raster documents (#3692)
* cargo fmt
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>
* fix(algo): mesh warps filter their source when shrinking, so thin lines don't alias
warp_triangles (Edit > Transform > Warp, smart objects placed through a warp, Puppet, Lens and
Wide Angle corrections) took one Catmull-Rom sample per output pixel with no minification
filter. A large smart object placed small through a warp (a 4200x4800 design at ~1/3-1/4 size
in a shirt mockup) came out with crinkled, doubled or missing grid lines where Photoshop's render
is smooth. The affine path (transform::warp_surface) already pre-reduces below 50% for exactly
this reason ("bicubic alone aliases below ~50%"); the mesh path had no equivalent.
Each triangle now carries its inverse-map Jacobian, computed analytically from its vertices.
Where it shrinks the source, an output pixel averages nx x ny bilinear subsamples over its
source footprint, with about one subsample per source pixel per axis (anisotropic, capped at 8).
Averaging happens in premultiplied space, before alpha is divided out. Triangles that keep or
enlarge the source still take the exact single sample, so identity and unscaled warps are
bit-for-bit unchanged. The per-tile source read grows by the subsamples' actual spread instead
of a fixed pad. Per-pixel filtering is used rather than a global pre-reduce because a mesh
warp's scale varies across the surface: one reduction factor over-blurs the parts that stay
large or under-filters the parts that shrink most.
Measured with an earlier cut of this approach (square bicubic subsampling) against Photoshop
exports of real mockups (RMSE of the placed design): 1.78% -> 0.81%
(White shirt), 0.49% -> 0.29% (Pepper). A mockup without a warp is unchanged.
Test: a_shrinking_mesh_warp_is_prefiltered (8/16/32-bit). A 1 px grid every 4 px placed at
1/5 through an Arc must stay within 0.25 of its 9/16 area average. Before: 0.5625 off (solid
black pixels). After: 0.20 off, which is the limit of a ~5 px box over a 4 px period.
Perf (30 MP RGBA8 source, place_source with Arc 10, release, M-series, best of 7):
1/3 size: 55 ms -> 87 ms
1/1 size: 240 ms -> 239 ms (no subsampling: identical path)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Use as_chunks for the opaque-pixel scan (clippy 1.99)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Free Transform: layer styles render with the transformed layer (#3653), work in progress
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Free Transform: a layer's style shows while the box is open, drawn from the transformed layer (#3653)
The live preview drew the moving pixels as a bare texture on a mesh, so Stroke,
Drop Shadow and every other effect vanished while the box was open (and for
type, shapes, groups and smart objects the style was baked into the texture,
cut off at the layer's bounds and stretched with the box).
A styled layer is now composited for the preview: the moving texels are
resampled through the box and the style is laid over them by the CPU
compositor's composite_with_effects, the code that draws the layer once the
transform is applied. A stroke keeps its width when the box is enlarged and a
shadow keeps its direction when it is turned, as after the commit.
The composite is made at screen resolution (never finer than the document),
with the style's sizes scaled to match, within a pixel budget from a measured
cost per effect: about 30 ms per rebuild while dragging and 120 ms at rest.
It is rebuilt only when the box, the zoom or the layer changes.
Still plain: the selected part of a styled layer, and Warp mode.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Merge fix: Warp on type reads the texels as the session did (bare when the box draws the style)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Format the type warp texel read
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Free Transform: Content-Aware Scale's preview has no styled composite
Main added a TransformPreview initializer for Content-Aware Scale (#3103) after this
branch started; give it the new field so the merge builds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Gradient Map: edit color stops on the ramp (#2958)
* Gradient Map: keep stop removal scoped to the editor
* Gradient Map: fit selected stop controls in narrow Properties panels
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A Displace smart filter (Fltr class `Dspl`, filterID 1148416108) was not in
the modelled filter list, so it became psd.unsupportedFilter and rendered
as a pass-through. Mockup PSDs that bend a replaceable design over fabric
(a design smart object with Displace then Gaussian Blur) rendered flat.
Photoshop stores the displacement map inside the filter: `DspD` (raw data)
holds u32 LE height, u32 LE width, u16 LE mode 3, then two 8-bit planes
(horizontal, vertical) and padding, next to HrzS/VrtS (scale %), DspM
(StrF stretch / Tile), UndA (RptE repeat / WrpA wrap), EmbF and DspF (an
alias to the author's map file, usually unavailable).
The filter now maps to filter.distort.displace. As for Camera Raw, the
whole item (with the map) is kept as hex in `__displacePsd`, so it
re-renders, survives .pcraft and is written back to PSD with only the
editable settings and blending replaced. params_for resolves the embedded
map, so the smart filter re-render applies it, stretched over the document
canvas (the bounds smart filters already pass). Another mode, a missing
or truncated map, or unknown enum values keep the item opaque as before.
The map parser uses checked sizes and caps maps at 2^26 pixels.
Tests: displace_map unit tests (layout, every truncation, hostile and
corrupted headers, round trip, edited settings, opaque fallbacks, bad
exports) and smart_psd::displace_psd_filter_renders_with_its_embedded_map
(8/16/32-bit: renders displaced, PSD export/import and .pcraft keep it).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A placed layer with a custom warp (warpCustom, a 4 x 4 customEnvelopeWarp
mesh) can have control points outside its source rect 0..w x 0..h. Photoshop
maps the bounding box of those 16 control points onto the Trnf quad. PhotoCraft
mapped the source rect onto it, so a mesh that spills past the source counted
the overhang twice and the re-rendered smart object came out stretched.
Measured against Photoshop's own export of a real shirt mockup (4200 x 4800
source, mesh spilling to x = -9, y = -88): before, a ~1.6% vertical stretch,
the tile offset growing from 0 to -16 px top to bottom; after, alignment is
exact everywhere and the RMSE against Photoshop is 0.49%. The bounding box of
the evaluated warp outline (rather than the control points) fixed that file
but put another at 5.4% RMSE, so it is the control-point box.
- blocks::placed_frame: the shared rule. The control-point box of a
single-patch custom warp, else the source rect; a degenerate or non-finite
box falls back to the source rect.
- blocks::parse_smart (and so smart_map::parse_sold): the affine maps that
frame onto Trnf.
- smart_map::quad_points: export writes Trnf (and PlLd's quad) as the image of
the same frame, so a spilled mesh round-trips and quad_moved keeps the
imported Trnf as stored. Distort / Perspective placements keep the source
rect, matching parse_smart_perspective.
Refs #579.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A Displace applied to a smart object is recorded as a smart filter and
rendered by apply_smart_filters -> filters::apply_filter_to_surface, both
when it is added and on every later re-render (replace contents, toggling
a filter, transforms). That path built FilterParams with params_for alone,
which leaves Displace's map as None; only run_filter resolved `mapPath`
(through filters_ext::resolve, which needs the session). With no map,
algo's displace returns the source unchanged, so the smart filter was a
silent no-op.
apply_filter_to_surface now resolves `mapPath` through the same file read
and import the direct command uses (filters_ext::map_from_path, shared by
resolve and the new session-free resolve_map_path). A map that can no
longer be read on re-render leaves the pixels alone, as for an unknown
filter; the command still refuses an unreadable map when the filter is
added. `mapDocument` needs the session and stays unresolved on re-render.
Regression test: smart_cmds::tests::
displace_smart_filter_reads_its_map_path_on_re_render (8/16/32-bit).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Diego Chagas
Jakub
Scott C Hughes
MrCloudy2andClaude Opus 5.5
Desayuno64
Reynold Hu
David MannandClaude Opus 5.5
UnemarsandClaude Opus 5.5
DennisandClaude Opus 5.5
JGFandClaude Opus 5.5
Valentin KleinandClaude Opus 5.5
Nicolas Falesy
Streetblock
