Session::bump took the disk cache's footage salt at every edit, which reads the size and time of
every footage file, each frame of an image sequence included. With three 5000-frame sequences a
caret move or a typed character in a text field took ~0.6 s on the UI thread (9 ms with the disk
cache off). The salt is now taken again only when the footage items changed, or when a project
is opened, footage is reloaded or the disk cache attached.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review: styled() kept the buffers it rendered instead of the ones LayerCache::insert returned, so
without a blending-space conversion concurrent misses still returned separate body and pass
buffers (and the GPU uploaded each).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Alt/Option-drag a Project item onto a Timeline layer, or press Ctrl+Alt+/ (Cmd+Opt+/) with layers and an item selected, to swap the layer's source. The layer keeps its name, timing, transforms, keyframes, expressions, effects, masks, switches, parenting and track matte; a source with audio adds Audio properties when missing. Precompositions can't be nested inside themselves.
AVLayer.replaceSource now runs the same engine command and closes the replaceSource part of #813.
Preview workers that start together miss the layer cache together, and each kept its own copy
of every layer; on the GPU each copy was uploaded and packed into its own Classic 3D atlas, which
ran the device out of memory on a cold start. LayerCache::insert now keeps the first buffer of a
key and returns it, and the renderer goes on with the returned buffer, so the workers share one.
Renders don't wait for each other: on the shared frame pool a blocking wait could end up beneath
the work it waits for (tried, it stalled playback for the wait's timeout).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Looking up to the whole RAM preview for uncached frames let the workers spread over the preview
while the playhead waited. On the GPU each worker then built a Classic 3D atlas for its own set
of layers, and on a cold start of the #1812 project the device ran out of memory far more often
(first pass 1-56 % of real time instead of 63-81 %). The window now keeps its old width but
starts at the first frame ahead that isn't cached, so a loop or a restart still renders on at
once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review: two footage items of one file with different nominal sizes could share a sized entry,
though one is resampled and the other kept as decoded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Edit > Copy with keys or Transform properties selected puts the After Effects keyframe text on the system clipboard, for other applications such as AE2Blend, in place of the 'EffectCraft: N keyframes' note. Ctrl+V in EffectCraft still pastes from its own clipboard.
Copying keys or unanimated properties of one layer records them as After Effects keyframe text: its Transform properties, one row per key with the frame in comp time (start offset and stretch applied) or a frameless row for a static value. EffectCraft's own keyframe clipboard is unchanged.
Review: mask motion blur draws an animated mask at the shutter's other times, but the layer key
hashed its value at this time only. Two frames with the same mask value (a hold before a move)
shared a cached layer, with the wrong blur. This applied to solids, text and shapes already, and
to stills since this PR.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review: still_at_size returned a smart object's pixels unresampled when they differed from the
footage's nominal size. It now follows footage_buf: pixels larger than nominal are resampled to
the requested size, smaller ones are left for the renderer to scale up.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Index-units test now opens the Units popup in the Timeline and picks
Index, asserts the stored Units value, then reads the painted Start, End
and Offset, so it also fails when the Units control doesn't update the
session.
Review: frame_at_size at a still's native size made a second cache key, so the pool decoded
and kept a second full-size copy of a frame thumbnail() or frame() already had.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The playback prefetch only looked ahead*3 frames (96 with 16 workers) for uncached frames.
After a loop or a restart from the start those were all cached, so the workers stood idle until
the playhead came within that window of the end of the cached run, then the preview stalled.
A playing preview now looks as far ahead as the RAM preview holds (and the range spans) for
the next uncached frames, still queuing at most ahead at a time; eviction already spares the
frames the playhead reaches first (#595).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Footage layers had no layer cache key, so a still made a new layer buffer every frame on every
preview worker, and the GPU compositor uploaded it again each time (its upload cache is keyed
by buffer). Stills (without a proxy) are now keyed like text and shape layers, without the
source time; movies and sequences are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The media pool honoured the requested size only for movies; stills came back
at full size and the renderer resampled them on every frame. A 21120 px wide
plate at Quarter was re-filtered each frame, and four of them (1.85 GB as f32)
did not fit the 1 GiB pool, so they were also decoded again and again.
Stills are now resampled once with effectcraft_raster::resample (the same
call the renderer made, so pixels are unchanged) and the pool keeps the
reduced copy under the sized key.
#1812 project at Quarter, CLI bench: 410 ms -> 33 ms a frame (26 ms with the
layer cache); desktop playback 2% -> 60-99% of real time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Illustrator EPS imported as an empty layer. Three gaps in the PostScript
interpreter, each against the PostScript Language Reference:
- An operator stored under another name (`/mo /moveto load def`, as
Illustrator's prolog does for every path and paint operator) pushed the
operator's name instead of running it, so nothing was drawn.
- Device queries that return values (currentpagedevice, currentscreen,
currentflat, currenthalftone, status, dictstack, ...) pushed nothing,
shifting the stack under the procset's guards. They now push a plain
device's defaults; setscreen pops its three operands.
- `currentfile 0 (eod) /SubFileDecode filter` now skips to the end-of-data
string, so an embedded XMP packet isn't run as PostScript.
The regression test uses synthetic PostScript only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The crop kept 3 px of halo, which makes the blur exact inside a box, but the
lag-1 correlation also reads the right neighbour's high-pass residual — one
column outside the box, still polluted by the crop edge when another box
covers it (adjacent boxes are the norm with a step of size/2). The size
estimate drifted by ~5e-5 in the parity test; a 4 px halo removes it.
The parity test now holds std to 1e-6 and size to 1e-7 (was 5 % for size),
and the boxed statistics still cost 1.8 / 3.1 ms at 1080p / 4K.
The sample-box search rescanned every candidate box and the grain statistics
blurred the whole image three times per layer, although Match Grain only reads
the 80 sample boxes (~6.5k pixels). On a 4K frame that analysis was ~90 % of
the effect.
- measure the statistics from small crops around each box (3 px halo, the
Gaussian's support), deduplicating overlapping boxes so the sums match the
previous masked version;
- search the candidate box rows in parallel (the sort is a total order, so the
chosen boxes are unchanged);
- compute the "own" boxes and stats only when Compensate for Existing Noise is
on, and clone the frame for Preview composition only in Preview mode;
- run Plane::transpose/map_rows on the calling thread for small planes: the
rayon dispatch dominated per-box crops (240 blurs of ~15x15).
Release, Ryzen Threadripper 1920X, noisy source / flat target, defaults, best
of 4 (1080p / 4K):
- grain statistics: 151/463 ms -> 1.2/2.8 ms
- sample boxes (source): 43/179 ms -> 16/71 ms
- end to end: 266/1416 ms -> 99/465 ms (Compensate 50 %: 392/1830 -> 104/482)
Tests: the boxed statistics match the masked ones (overlapping and corner
boxes, alpha holes), plus the existing Match Grain/Remove Grain suites.
xtask bench passed the whole layers scenario on one command line, longer
than Windows' 32 767-character limit (os error 206), so no scenario past
demo ran. effectcraft-cli run takes --args-file F (one argument per line)
and bench writes its commands there.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- first_ascent: a "d" shorter than 0.3 x the font ascent (a symbol or icon font) no
longer puts box text's first baseline almost at the box top; the font ascent is used,
as before the "d" rule (CodeRabbit on #1875). Test:
a_short_d_falls_back_to_the_font_ascent.
- CharXf::track_before / track_after are pixels, not 1/1000 em (CodeRabbit).
- The Lottie export of style-run tracking is back to what it was: the pixel conversion
was not checked against a player, so it is left for a separate, verified change.
- The Subtract clamp in Mode::combine is intended for every animator property: the
selectors combine into one selection per character for the whole animator, as in
After Effects; the read-back observed it through Tracking (now said in the code).
- Per-character 3D rotation notes that #1876 adds Mat4::rotation_xyz for the same order.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
kernels.wgsl repeats Glow's bright-pass gain 8.0 and knee 0.0611 as literals; they now
name effects::GLOW_GAIN / GLOW_KNEE (exported) and the test
glow_constants_match_the_cpu (no GPU needed) fails if they drift. docs/effects-parity.md:
the summary row no longer says no effect is compared with After Effects while the
Blur & Sharpen row reports Radial Blur's comparison (CodeRabbit on #1873), and the
Radial Blur changelog row says CPU only (it has no GPU path).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WGSL cannot read Rust constants, so the Rec. 601 luma (Tint, Tritone, Luma Key /
Extract), Black & White's tint weights and Exposure's 2.2 power stay literals in the
kernels; they now name their Rust source (effects::LUMA_601, BW_TINT_LUMA,
EXPOSURE_LINEAR_GAMMA, newly exported) and the test shader_constants_match_the_cpu
(no GPU needed) fails if a literal drifts. gpu_tint_defaults_match_after_effects_greys
says "skipped: no GPU adapter" when it skips (EFFECTCRAFT_REQUIRE_GPU=1 still makes it
fail). docs/effects-parity.md: Brightness & Contrast is close to, not the same as,
After Effects (up to 6.2/255), and the summary row lists what is measured.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tylerjenningsw
dedandClaude Opus 5.5
Camille Moine
Brandon ThomasandClaude Opus 5.5
SBaksa
vangel76andClaude Sonnet 5.5
TimotheusG
Egon 5000
Wes CampbellandClaude Opus 5.5
NetterSilverandClaude Haiku 5.5
Davit MatchakhelidzeandClaude Opus 5.5