With Spacing unchecked, dabs are spaced by pointer speed, and a segment
without timestamps got a single dab at its end point. The canvas sends no
timestamps, so a Shift-click straight line with such a brush (the Eraser
with 10% smoothing gives start, 90% and end points) left three separate
dabs instead of a line. Untimed segments now fall back to distance spacing.
Co-authored-by: Jason Short <3344856+jason4short@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
tools.setBrush and stroke brush patches accepted any Scatter amount. The paint path
multiplied it by the radius in f32, so 1e38 overflowed the dab centre to infinity, and
dab_rect's unchecked i32 arithmetic then panicked (wrapped to an empty rect in release).
Brush validation now rejects primary and Dual Brush Scatter outside 0..=MAX_SCATTER (10,
the slider's 1000 %) with a BadParams naming the value, before the session brush changes
(also between setBrush's two merges, so infinity is named rather than reported as null).
bounded_for_render caps both amounts for direct paint API users (NaN scatters nothing),
and the dab rectangle, now one helper shared with the retouch tools, saturates.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
In the Photoshop themes (the default ProMedium included) the Brush and
Eraser options bar drew three icon buttons that did nothing when clicked.
The engine already supports everything behind them:
- Airbrush toggles the brush's Build-up (lit while on).
- The gear beside Smoothing opens Photoshop's smoothing options: Pulled
String Mode, Stroke Catch-up, Catch-up on Stroke End and Adjust for
Zoom. They edit the brush like the Smoothing % does, so they stay per
tool and reach the live stroke and the commit.
- The symmetry button opens a menu with Symmetry Off and every path the
Paths panel lists (saved paths, the Work Path, the selected layer's
shape path or vector mask), running paint.symmetryFromPath /
paint.symmetryDisable. It is lit while symmetry is on; with no path it
says how to make one.
The three buttons get accessible names (brush_picker::named), and
"Symmetry Off" is added to the twelve complete catalogs.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Median gathered every sample of the disc (about pi r^2 values) for each pixel and channel and selected the middle one, so the work grew with the square of the radius. For radius 2 and up, rank each channel's window samples by total_cmp (8-bit samples by level), keep a histogram of the disc's ranks that slides in a serpentine over the tile (one end of each disc row or column out, the other in, per step), and track the median rank as it moves (Huang), with block counts so it crosses large empty rank ranges a block at a time. The value picked is the sample the selection picks, so output is bit-identical at every depth. Radius 1 keeps the direct selection (now median_direct), which is also the test oracle.
layer.linkLayers picked the new link group as max(link_group) + 1, computed even
when unlinking. A .pcraft or PSD can store u64::MAX, so the addition panicked in
checked builds and wrapped to 0 in release, where it could merge the selection
into an existing group 0.
Add Document::unused_link_group: max + 1 when it fits, otherwise the smallest
free id from 1, never wrapping. Link Layers uses it only when linking and returns
an error if no id is left. Also read the first selected group and the member
counts without indexing.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The midpoint action checked `i + 1 >= v.len()` on the raw `index` param,
so an index of usize::MAX overflowed (a panic with overflow checks on, a
silent no-op success with them off). The shared index parser now takes
the number of addressable entries and rejects an index past them before
any arithmetic, for every action (move, delete, color, opacity, midpoint)
and both kinds, so all of them validate the same way.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Follow-up to #991: Rust 1.99's clippy flags chunks_exact with a constant
size, which turns the workspace clippy gate red.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
timeline.create and timeline.setProps kept any `duration` (up to u64::MAX) and
any `fps`, and timeline.nextFrame/previousFrame then moved the playhead with
`(current as i64 + delta).clamp(0, duration as i64 - 1)`, which panics when the
upper bound goes negative (duration 0 or above i64::MAX) or overflows.
Both commands now reject a `duration` outside 1..=1000000 and an `fps` that
isn't finite and in (0, 1000] with BadParams before anything changes. Stepping
moves to a saturating Timeline::step, safe on any stored timeline (e.g. one
loaded from a file with duration 0), and a non-finite stored fps falls back to
30 like a non-positive one.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The bar height was clamped to [2, h/4], and f64::clamp panics when h/4 < 2,
so any image shorter than 8 px panicked inside the command's run closure.
Reject such images with a bad-params error before anything changes, like the
existing bar-width check. Also keep the bar on the canvas when the label under
it, or the left margin, would push it off a small image.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
crop_doc negated the crop origin unchecked, so x or y = i32::MIN overflowed;
an origin one past i32::MIN pushed layer pixels past i32::MAX and panicked on
a length mismatch while translating surfaces; and x + width past i32::MAX
saturated, silently cropping less than asked. crop now checks the far edge,
crop_doc checks the negation, and translate_doc checks that every surface
still fits before moving anything; all return BadParams naming the values.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Surface::fill_rect materialized every tile it touched, so a new document's
background cost one 256x256 tile per tile slot: ~360 GB for the 300000x300000
size the command accepts. Fully covered tiles now share one Arc'd tile, so a
solid fill costs one tile plus the tile index; edge tiles are written as
before, and copy-on-write splits a shared tile off when it is edited.
Duplicating a layer (Layer › Duplicate, ⌥-drag with the Move tool, Layer
via Copy without a selection, New Smart Object via Copy, the Edit menu's
duplicate, Select and Mask's new layer) names the copy with
`Document::copy_name`: "Layer 1 copy" first, then "Layer 1 copy 2",
"Layer 1 copy 3"…, unique in the document. A name already ending in
"copy" or "copy N" is numbered from its root instead of growing another
"copy"; names that merely contain the word keep it.
Co-authored-by: Jason Short <3344856+jason4short@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When the clipboard holds an image (pixels copied in the app, or an image
copied in another app), File › New offers a Clipboard preset first under
the Recent blank-document presets, at that width and height and 72 ppi,
and selects it, so Create makes a document the size of the clipboard.
Without an image the dialog opens as before. Opening the dialog is an
explicit request, so the OS clipboard is read then, as for a paste.
Co-authored-by: Jason Short <3344856+jason4short@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
record_surface sized the interleave buffer from the layer rectangle alone,
so a few-hundred-byte file declaring a 300000x300000 layer with no
channels made import allocate ~360 GB and abort. The channel decoders
already bound their output by the bytes present, so a buffer is now built
only when at least one channel decodes; otherwise the layer imports empty
with a warning.
- Color Lookup: a stored document's `size` is untrusted. `size³ × 3` overflowed for sizes of
2²² and up: debug builds panicked, release builds wrapped past the table-length guard and
indexed out of bounds. One checked `photocraft_doc::adjust::lut3d_len` now backs the CPU
compositor, the GPU planner and the PSD writer.
- Quick Selection: a stroke point far off the canvas (`select.quick {"points":[[3e9,0]]}`)
saturated to i32::MAX and then overflowed on `+ 1`.
- Link Layers: `max + 1` over stored link group ids overflowed at u64::MAX (debug panic, release
wrap that could merge unrelated layers). A fresh id is now the smallest unused one when the
largest is taken; a selection's first group is read with `first()`.
- Vanishing Point clone footprint: the same `as i32 + 1` saturates (hardening; no reproducer).
Per maintainer review (#833): keep Feather as Difusão…, Diffuse becomes
Difundir…, and Liquify becomes Liquefazer… (Licuar is Spanish).
Separates the Fast menu and Stylize menu entries that were colliding in pt-BR.
The contributor docs offer github.com/storytold/craftrules as the place
to read the font rules and release playbook, but the repository
returns 404 to everyone outside the organization, so the cited text is
unreachable for outside contributors. Keep the sibling-checkout paths
and say plainly that the repo is not public and a maintainer can supply
the rules relevant to a change.
Co-authored-by: Brandon Fenty <bfenty@users.noreply.github.com>
Hovering a mode in the Layers panel's blend-mode list previews it on the
selected layer; moving off the list reverts, and only a click applies it
(one history step), as in Photoshop.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The margin Modify pads a canvas-edge selection with was twice the radius
for every operation, so Feather 1000 added 2002 px a side. It now pads by
each operation's reach (Feather 1.5 r, the others r) and refuses a padded
mask over 2^30 pixels instead of aborting on the allocation. The new
dialog option is translated in every shipped language.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UHgFEHEsefoQoBParH47JA
Contract, Smooth, Expand and Feather gain Photoshop's "Apply effect at canvas bounds" option
(applyAtCanvasBounds, default off). On, the canvas edge counts as a selection edge, so Select All
then Contract shrinks from the edges and Feather softens them; off, the selection continues past
the canvas, as in Photoshop. Border always follows the canvas edge, so a Select All borders to a
band inside it instead of an empty selection.
The mask is padded (empty or edge-replicated) only when the selection touches the canvas edge,
so selections clear of it cost nothing extra.
1000000 % on any layer could ask the transform for an image millions of
pixels wide, more than can be allocated. The W and H fields now reach as
far as the box stays within 300000 px (the document limit), at most
1000000 %, and a linked change keeps the other side within its limit too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UHgFEHEsefoQoBParH47JA
The options bar clamped W/H to ±10000% and X/Y to ±30000 px, while the engine takes any scale
and documents go up to 300000 px a side. Overlays on large stitched mosaics need more. The limits
are now named constants: ±1000000% (a 300 px layer can span the largest document) and ±300000 px.
File › Open, the web build, batch folders and the CLI now pick up
.afdesign, .afphoto and .afpub, which the native reader handles; the MCP
test checks the preview fallback's new warning.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GQPh8UDgMyARR5Hat9UchC
The Photopea screenshots, their ATTRIBUTION row and the example that made
them documented the bug for review; the unit tests in crates/text cover
the behaviour, so they are not needed in the tree.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UHgFEHEsefoQoBParH47JA
Dragging one live-gradient handle towards the other clamps the fill's scale at 10%. For Linear the
centre was still taken from the midpoint of the requested points, so the
clamped gradient was re-centred and the handle that wasn't being dragged slid along the line.
gradient.fill.set now shifts the clamped layout so the handle not given keeps its place. The same
apply_set drives the drag preview, so the canvas matches.
The Filter Gallery, Camera Raw and Adaptive Wide Angle dialogs record the layer they ran on in
the command's params. Last Filter replayed the journal entry verbatim, so it filtered that layer
again (possibly hidden, or deleted) and appeared to do nothing on the layer you were looking at.
It now drops the recorded `layer`, so it applies to the current layer as in Photoshop.
Distort and Perspective move a corner along the drag only as far as the
box stays convex, stopping just short of flat, as Photoshop and Photopea
do, instead of folding it into a bow tie.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Pro themes already had Photoshop's Auto-Select checkbox; the Studio
themes showed only a hint for the Move tool, so Auto-Select could not be
turned off there. Both now have the toggle.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Right-click with a selection tool and choose Delete and Fill Selection
to fill the selection from its surroundings with Content-Aware Fill's
default settings, without the dialog. Engine command
edit.deleteAndFillSelection, one undo step, translated in every shipped
language.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
photocraft-affinity now reads native Affinity documents (container
versions 8-12): the FAT and revision chain, stored/zlib/zstd entries
with predictors and CRC checks, the tagged object stream in doc.dat,
and the document model (pages, artboards, layers, groups, curves,
geometric and compound shapes, fills, gradients, strokes, text, placed
images, pixel layers, masks and embedded-document caches). Every size
is bounded and malformed input returns an error.
photocraft-io maps it to an RGB8 document with artboards, groups, shape
layers, type layers, embedded smart objects, pixel layers and
vector/layer masks; what can't be mapped is listed in the warnings. The
embedded preview is now only a fallback when the native data can't be
read, and only that preview is refused by Place and warningless imports.
`cargo xtask corpus --affinity` pins 21 public CC0/MIT documents,
sha256-verified; the io corpus test compares each import with the
thumbnail Affinity saved in it (mean difference 0-4.6 of 255).
Affinity export is not implemented: no Affinity installation was
available to validate written files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GQPh8UDgMyARR5Hat9UchC
* Allow removing last polygonal lasso vertex without editing layers
* Route lasso Backspace ahead of destructive edit shortcuts
* Use right-click to retract polygonal lasso points instead of opening menu
* Regress lasso vertex removal via keyboard and actual canvas right click
* Apply rustfmt layout reported by Linux CI
* Apply rustfmt layout reported by Linux CI
* Remember last applied New Guide Layout parameters in UI state
* Save guide layout preferences only after successful application
* Cover successful, cancelled and failed guide layout dialog state
* Avoid overlapping borrows when reopening guide layout dialog
The document parameter's set_active result was discarded, so an
out-of-range index still switched nothing, changed nothing, and the
window opened for whatever document happened to be active - a caller
was told success for a document that does not exist. Now the call
errors with the index and leaves the active document alone.
real_before_parameters picked Photoshop's order (real mask first) whenever
the record length fit the parameter-flags byte at offset 36, even when that
byte was just part of the real rectangle. A density-only record in the
spec's order then read its density as the real mask's background and lost
the parameter. When both orders fit, Photoshop's reading now has to be
plausible (background 0 or 255, known parameter flag bits) or the spec's
reading with a real mask wins.
Co-authored-by: wstock <1394687+wstock@users.noreply.github.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
apply_to_doc replaced a text-replacement variable's text and set the layer's
render cache to None, but nothing rebuilt it. The compositor draws a type
layer from that cache, so the text vanished from the canvas after
image.applyDataSet and from every file written by
file.export.dataSetsAsFiles. The layer is now re-rendered with
type_cmds::refresh, as type.edit does.
Motion Blur took distance + 1 bilinear samples per pixel, each through Image::sample with per-tap edge checks, so distance 100 cost about 400 edge-checked reads per pixel and channel. The sample offsets are the same for every pixel, so merge their bilinear taps once into a fixed kernel (each tap an integer offset and a weight) and apply it tap by tap as a multiply-add over whole rows of a premultiplied window, which the compiler vectorizes. This is the same premultiplied mean as before, summed in a different order.