[vanceingalls]

What

Motion blur's filter tier now follows the host's preview quality. In preview it draws 32 offset copies per mover at Full, 16 at 1/2, and 8 at 1/4 and 1/8. Renders (engineMode) always draw 32. The level arrives through the existing set-preview-quality bridge message, which already drives the lighter video copies.

The runtime's preview facts also change. lighterAt used to list only levels where a heavy video gets a lighter copy. It now also lists 1/2 and 1/4 once the filter tier is on and has movers. Hosts such as HyperFrames Desktop step quality down on choppy playback only to a level listed in lighterAt. Without this, a film whose only load is blur was never stepped down (no_lighter).

Why

On GPU Chrome, preview lag with blur on comes from the GPU process rasterizing each mover's SVG filter: 32 feOffset and 31 arithmetic feComposite. Cost grows with the number of movers moving at once. On a 60 Hz headless Chrome 157 (M4 Pro), 3 interleaved reps, fps with blur off, then 32 / 16 / 8 copies:

• 50 movers: 59.5, then 17.5 / 29.6 / 52.1.
• 100 movers: 59.9, then 6.9 / 11.8 / 25.6.
• k3-promo (106 movers): 59.6, then 42.8 / 43.8 / 52.9.

The machine was loaded (load average 45 to 92), so only the order of the arms is quotable, not the absolute fps.

Picture quality was scored against a sub-frame reference: the blur-off film averaged over 32 instants across the shutter. The measure is the engine oracle's worst 16 px tile; thresholds.ts puts the smallest real symptom at 14.47.

• 16 copies against 32: worst tile 11.9 or less on every frame measured.
• 8 copies: passes on translating content, but reaches 28.5 on turning content in k3-promo. That is why 8 waits for 1/4.
• A single feGaussianBlur was also tested and rejected: worse on every kind of motion, and it still costs GPU time for each mover.

Preview and render

At 1/2 and below, preview and render differ by the numbers above. The motion blur guide now says so. The capture tier is unchanged.

Tests

• One new case in captureMotionBlur/index.test.ts: 16 copies at 1/2, 8 at 1/8, 32 for an engine frame, and lighterAt levels only once movers are found. It fails if the preview tap count is reverted.
• All 243 captureMotionBlur tests pass.
• Core typecheck, oxlint and oxfmt pass.

Rollout

Desktop picks this up after the next OSS release and a pin bump. No Desktop code change is needed: the quality row and the auto step-down already send the level.

🤖 Generated with Claude Code

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
hyperframes	🟢 Ready	View Preview	Nov 11, 2026, 8:02 AM

💡 Tip: Enable Automations to automatically generate PRs for you.

[github-actions]

Edit accuracy: accurate 267 (base branch 267), smooth 112 of those

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (8)

• seqplainkeys-none-center-r0-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): pass, passes
• seqplainkeys-none-center-r0-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): fail / pass / pass, passes
• seqplainkeys-none-center-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-pct-r0-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): pass, passes
• seqplainkeys-none-pct-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-px-r0-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): fail / pass / pass, passes
• seqplainkeys-none-px-r30-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-px-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run

[heygengenesis]

Verdict: APPROVE · 0 blocker · 0 important · 0 nits · head 961dfb6

Preview taps follow the level, renders stay at 32 (engineMode), and planFilter's step-sized softening scales with the wider gaps between copies on its own. CI is green on this head.

Scope

• Read all 5 files end to end, plus filterPlan.ts, frames.ts/tapsFromReads, mediaProxy.setPreviewQuality, and the Desktop step-down consumer (QualityToast.tsx onChoppy → lighterBelow).
• Rebuilding the layer in place (layerOf removes the old <filter> and builds a new one under the same id): I checked this in Chromium headless-shell 1237. An element styled filter: url(#hf-mb-0) picks up the replacement filter on the next frame without its style being touched.
• lighterAt: Desktop only asks for facts when playback turns choppy. By then a paint has run discovery, so the movers > 0 gate is satisfied. At 1/4 only videos can list 1/8, which matches 8 copies at both levels.
• One thing to watch after the pin bump: the step-down still falls back to busy when long main-thread frames reach 50% or more. On a film with many movers, the filter tier's per-frame pose reads count toward that share, so check Desktop's stepDownDeclined() on a blur-heavy film.

— Genesis