Anima on the web

apps/anima-web runs the same engine and egui UI as the desktop app in the browser: the workspace compiled to wasm32-unknown-unknown, started by eframe's web runner on WebGPU (WebGL2 fallback). It is a static site: one .wasm, its wasm-bindgen JavaScript glue, index.html, a few small scripts (worker, audio worklet, service worker), a web manifest and icons. Nothing is uploaded anywhere: projects, media, settings and renders stay in the browser (its private storage) and on the user's machine.

Build and run

cargo xtask web                 # release build → <target>/web/dist
cargo xtask web --dev           # unoptimised build (faster to compile, slow to run)
cargo xtask web --serve 8765    # build, then serve dist on http://127.0.0.1:8765/

<target> is $CARGO_TARGET_DIR or target. dist holds index.html, anima_web.js + anima_web_bg.wasm (+ snippets/, the browser glue js/host.js), worker.js, audio-worklet.js, sw.js (its precache list and version filled in by the build), manifest.webmanifest, favicon.svg and icon-256.png / icon-512.png (copied from assets/app-icon). --serve is a tiny localhost-only static server that sends Cross-Origin-Opener-Policy / Cross-Origin-Embedder-Policy, so the page is cross-origin isolated like a production deployment should be. Any static server works (python3 -m http.server -d target/web/dist 8765); it must serve .wasm as application/wasm. Service workers need a secure context: https://, or http://localhost / 127.0.0.1.

The release .wasm is about 58 MB before wasm-opt (v0.3.1; code: the effects, codecs, the expression engine and the UI; data, mostly the bundled fonts; function names for readable panics). Served compressed it is about 17 MB with gzip or 10 MB with Brotli, and the service worker caches it after the first visit. Its name doesn't change between releases, so hosts revalidate it rather than caching it for long (packaging/web/README.md).

On other targets the web crate holds only its portable storage model (store.rs, unit-tested by cargo test --workspace); cargo xtask wasm (part of cargo xtask ci) checks the whole crate, with every L0–L4 crate and anima-ui-egui, for wasm32-unknown-unknown.

URL flags: ?empty (a blank project; the session is not restored or recorded), ?demo (the demo project instead of the last session), ?home (show the start screen), ?storage=indexeddb / ?storage=memory (force a storage backend), ?noworkers (render, analyse and draw viewer frames on the page's thread), ?frameworkers=N (frame workers, default 2; 0 = viewer frames on the page's thread), ?nogpuworkers (frame and job workers render on their CPU, without a WebGPU device of their own), ?nosw (don't register the service worker).

How the desktop pieces map to the browser

Persistence

Everything that lives in files on the desktop lives in the browser's Origin Private File System (OPFS); where OPFS can't be written from the page (older Safari), IndexedDB; with neither (some private windows), memory only. anima.info().storage.backend says which. The page asks for persistent storage (navigator.storage.persist()), which browsers grant to installed apps, so the store isn't evicted under storage pressure.

The engine's storage traits are synchronous and the browser's are not, so the app works on an in-memory mirror (store.rs): every stored entry is read before the app starts, reads and writes hit the mirror, and changes flush in the background (one write per changed key, in order; OPFS writes are atomic: a swap file replaced on close). Keys:

On the next visit the app reopens the snapshot (unsaved changes stay unsaved: the project is modified until saved, Session::mark_unsaved, so File ▸ New, Open, Close Project and Revert still ask to save it first), registers every stored media file with the media pool so its footage decodes, and File ▸ Open Recent reopens saved projects from storage. Without a snapshot (first visit) it starts with a blank project (?demo opens the demo project). Render outputs are not stored: they download.

Background jobs: Web Workers

wasm threads need shared memory, which needs a nightly build-std toolchain, so this build is single-threaded per instance. Long jobs run in Web Workers instead, each with its own instance of the same module (the page compiles it once and posts the WebAssembly.Module, so nothing is compiled twice) and its own memory (crates/engine/src/offload.rs, src/worker.rs):

1. The session's Offload serializes a WorkerRequest: the project JSON, the footage paths it reads, and the job (render with the resolved queue items and output paths; warp, camera, track, maskTrack, rotoFreeze, rotoPropagate with the target and analysis parameters; contentFill with its plan).
2. js/host.js takes an idle worker from its pool (or starts one), sends it the footage bytes it doesn't have yet, then the request.
3. The worker runs the job on a plain engine session and posts WorkerReplys: render progress (at most every 100 ms), item started/finished, rendered files (transferred), analysis progress, the analysed property group, failed, done; then {type: "stats"} with the job's GPU passes and readbacks. With a WebGPU device of its own the job's frames render on it in passes (below, GPU in job workers).
4. The page applies them every frame: Session::poll_render updates the queue items and the progress bar, Session::poll_offload writes an analysis result as one undo step (the effect / tracker / mask group is replaced; other edits made meanwhile are kept) and the downloads start.

Stopping a job terminates its worker (a worker can't be interrupted mid-frame): the item being rendered becomes "User Stopped", an analysis writes nothing. Every job shows in Window ▸ Progress (jobs.list) with its progress and a cancel button (jobs.cancel).

Roto Brush propagation runs in a job worker too (WorkerJob::RotoPropagate): the worker streams the segmentations it computes back at most every 250 ms (WorkerReply::Segs: each FrameSeg in a compact binary form, run-length coded planes, base64), the page puts them into its segmentation cache and relays them to the frame workers. Stopping it keeps the frames propagated so far.

wait: true (the default of renderQueue.render for agents, and an option of the analysis and Roto Brush commands) no longer blocks the page: over window.anima (the control channel) such a command runs with wait: false, and the promise resolves when the job it started ends, with what the blocking command returns (final render item statuses, the analysis status) or rejects with the job's error (ui-egui/src/control.rs, JobWaiter). Engine commands executed directly on the session (scripts) still block.

Viewer frames: frame workers

Viewer frames (the frame on screen, RAM preview and prefetch) render in long-lived frame workers (src/frames.rs, crates/engine/src/remote.rs), two by default. Each holds a replica of the project:

1. Before a render, the page sends a worker the footage files it lacks, then a sync: the whole project the first time, afterwards a diff of the serialized project from the revision the worker holds (remote::diff: object members added / changed / removed, arrays element-wise when their length is unchanged; a slider drag sends a few hundred bytes, not the project).
2. Then the render request (FrameMsg::Render: revision, comp, time, render options). The worker renders on its CPU renderer and posts the frame back as premultiplied RGBA8 in a transferred buffer (FrameReply::Frame); the page puts it in the frame cache.
3. One frame per worker at a time, so the page always picks what renders next (the viewer's frame before prefetch); a request for a revision the worker does not hold is refused, a worker that lost its replica asks for the whole project again (FrameReply::Resync).

Frames without effects that the WebGPU compositor can finish alone still render on the page (they stay on the GPU, no readback); everything else (frames with effects, 3D runs, the Software Only renderer, WebGL2) goes to the workers. If no worker is usable (one that failed to start, or one the browser refused to create, as in a sandboxed iframe), frames render on the page's thread as before. anima.info().frameWorkers reports {workers, alive, busy, rendered, failed, lastMs, syncs, syncBytes, gpu, gpuFrames, lastPasses, diskHits} (gpu: each worker's adapter, {cpu: why} without one, null until it reported).

GPU in job workers (M13.30). Each job worker opens its own WebGPU device too (workerJobInit; ?nogpuworkers: their CPU) and makes it the session's accelerator. A job can't block on a readback either, so the job loops are futures rather than blocking loops: anima_engine::offload::run_request_async runs the request, and every frame it renders goes through anima_render::passes (in_passes, comp_frame): render a pass, and while it missed, await the device (Accelerator::settle, the browser's event loop delivers the readbacks) and render again, as FrameServer::resume does for viewer frames (the desktop drives the same futures with passes::block_on; they never wait there).

• Render Queue: anima_export::export_async renders each frame in passes (frames with deferred readbacks render one by one instead of in parallel batches) and encodes as before (Exporter::export_async). As on the desktop, a job renders every frame on the project's renderer (the worker's GPU with Mercury GPU Acceleration), not on whichever compositor measured faster, so runs don't differ by timing.
• Analyses (Warp Stabilizer, 3D Camera Tracker, Track Motion, mask and face tracking, Roto Brush freeze and propagation): their work functions are async; started with wait inside a job they are queued and awaited (offload::run_or_queue), and each input frame renders in passes with the GPU backend (offload::job_accel). Roto Brush's segmentation reads frames through a callback, so its frame source prefetches the frame and its neighbours in passes before each step. On the desktop analyses render their input on the CPU, as before.
• Particles and Advanced 3D read back under keys like effect chains (crate::deferred): a particle simulation by its request (system, steps, births), a rasterised scene by everything it uploads, a whole prepared run by its scenes and output; the first pass starts the readback and gets a placeholder (no particles, an empty target), later passes read the bytes. A run with depth of field resolves on the CPU (its blur needs a readback mid-run), each of its scenes still rasterised on the GPU. So frames with particles or Advanced 3D render on the device in frame workers and job workers alike.

anima.info().workers reports {running, idle, gpu: {adapter} | {cpu: why}, jobs: [{kind, gpu, passes, readbacks, ms}]} (the last jobs).

Warp Stabilizer plans off the page (M13.32). Solving a stabilization plan (Subspace Warp: an eigen-decomposition per frame, then a mesh fit) takes seconds for a long clip, and warp.status used to solve it on the page the first time it saw the analysis (a ≈ 5 s stall). Now the analysis job solves the plan for the analysed settings and sends its summary (WorkerReply::WarpPlan: per-frame warps, auto-scale, crop, valid fraction) with the result. When the settings change later, warp.status reports stabilizing: true and starts a WorkerJob::WarpPlan job (jobs.list id warpPlan) instead of solving it; the next status after it ends has the plan. Frames with the effect render in the frame workers, which solve their own plans. The desktop solves the plan in place, as before.

Content-Aware Fill in a job worker (M13.30). contentFill.generate (the panel's Generate Fill Layer) sends its plan (content_fill::FillPlan: layer, frames, method, folder) to a job worker (WorkerJob::ContentFill, jobs.list id contentFill): the worker renders the layer's frames in passes on its device, fills them and writes the PNG sequence through its services, which send each file to the page's browser storage ({type: "store"} messages, under /Fill/<layer>_Fill_<n>/); then the plan comes back (WorkerReply::Fill) and the page adds the fill layer as one undo step. (Before, the fill ran on the page's thread and could not write its folder.)

Threading design: one engine instance per worker (decided, M13.24). Shared-memory threads inside one instance (SharedArrayBuffer + atomics, a wasm-bindgen-rayon-style pool) need the standard library rebuilt with atomics (-Z build-std, nightly Rust) and every crate's statics made thread-safe for the browser's rules (no blocking on the page's thread). The repository builds on stable Rust, so concurrency comes from workers that each run their own engine instance with their own memory: job workers (renders, analyses), frame workers (viewer frames, each with its own GPU device) and the page. What crosses between them is explicit: project diffs, footage bytes, frames, Roto Brush segmentations, cache keys. The cost is memory per worker (a project replica, the footage it reads, a layer cache) and serial loops inside one frame (rayon's pool falls back to the calling thread). The page is still cross-origin isolated (--serve and the service worker send COOP / COEP), so a threaded build (build-std with +atomics,+bulk-memory, wasm-bindgen-rayon) could be added later without changing the deployment; it would parallelise the loops inside a frame, not replace the workers.

Media Browser

Window ▸ Media Browser browses a virtual tree (src/browse.rs):

• Browser Storage: every file in browser storage (imported media, saved projects, files added with Add Files… or Upload Folder…, plain file inputs that work in every browser). This is the persistent recent list where folders can't be opened. A picked file that can't be read is named in the status bar with the browser's reason, and the others are still added.
• Folders: Open Folder… (File System Access API, showDirectoryPicker, Chromium) lists a folder from the user's disk; only the listing is read, and a file's bytes are read when it is imported (then kept in browser storage). The folder handles are remembered in IndexedDB; after a reload the folders come back if the browser still grants access, else Reconnect asks again.

mediaBrowser.list reports the places and these actions; mediaBrowser.action {action} runs one.

Audio

Audio preview goes through Web Audio: an AudioWorklet (audio-worklet.js; a ScriptProcessorNode where worklets are unavailable) plays interleaved stereo blocks. The mixdown is the desktop's (AudioPlayback, mix_comp): without a feeder thread, AudioPlayback::pump mixes ahead on every UI frame (≈ 350 ms queued) and the output device takes what keeps it ≈ 150 ms ahead of the playhead. A/V sync follows AudioContext.currentTime: the playback clock is the frames played by the context clock minus its output latency, and the viewer shows the frame under the audible sample, as on the desktop. The Audio panel's meters read the same feed.

Browsers start an AudioContext suspended until the user interacts with the page; the app resumes it on the first pointer or key event, so audio plays from the first preview after any click or key press.

GPU

eframe starts on WebGPU where the browser has it (WebGL2 otherwise), and the GPU compositor (anima-gpu) runs on the same device: viewer frames are composited by compute shaders and drawn straight from their texture, with no readback. The Info panel's pixel readout and the eyedroppers read GPU frames back asynchronously (Gpu::read_display_async: the pixels arrive a frame later). On WebGL2, or without a usable adapter, everything renders on the CPU. anima.info().gpu reports {compositor, viewerOnGpu}.

If the page's WebGPU device is lost (a driver reset, the browser reclaiming the GPU), the canvas can't show anything new, while the project stays open and commands keep working. The page then covers the canvas with a notice (showDeviceLost in js/host.js) offering Save Project (a download) and Reload, which first writes pending changes to browser storage (anima.flush()) so the session comes back after the reload; when the write fails it says so and offers Reload Anyway.

GPU effects run in the frame workers. A browser never lets JavaScript (or wasm) wait for a GPU readback, not even in a worker: a buffer maps only once the thread returns to its event loop. The CPU renderer, which runs GPU effect chains and reads their results back mid-frame, therefore can't use the page's device. Instead each frame worker opens its own WebGPU device (navigator.gpu in dedicated workers: Chromium, Safari Technology Preview; Gpu::request_deferred) and renders with deferred readbacks (crates/gpu/src/deferred.rs; job workers do the same, above):

1. Every readback the renderer needs is keyed by what determines its pixels: a GPU effect chain by its input buffer's pixels and the chain's parameters, the top-level frame by comp, time and render options.
2. A key seen for the first time records the GPU work, starts an asynchronous copy (map_async) and misses: the step returns a cheap placeholder (its input), and the pass is marked missed. From the first miss to the end of the pass, the layer cache drops inserts (LayerCache::set_gate), so no placeholder-derived buffer is kept.
3. The worker awaits the device (Gpu::settled) and renders the frame again (FrameServer::resume); keys whose bytes arrived are served from them. A pass without a miss is the frame. Chains that depend on other chains (a precomp with effects inside, effects that read other layers) take one more pass per level; independent chains resolve together. A chain whose kernels read data that missed while they ran (through the effect host) is not started, so a result is only ever computed from genuine inputs. After 24 passes a frame renders on the worker's CPU.

The whole frame renders on the GPU (the compositor walk with GPU effects, then one readback) when it can; steps that would need a mid-walk readback (an Advanced 3D run the kernels decline, CPU-only effects on an adjustment layer) make the frame composite on the CPU instead, with its GPU effect chains still on the GPU. Backend Auto works in the workers: it picks CPU or GPU per comp from the frames' total times across their passes (AutoPick), as on the desktop. Without WebGPU in workers (Firefox, older Safari) or with ?nogpuworkers, the workers render on their CPU, as before. A frame with effects goes to a worker when the workers have a GPU, so the page never runs an effect on its thread. Particles and Advanced 3D read back under keys too (M13.30, above).

anima.workerFrameCheck({time?, scale?, backend?}) renders the active comp's frame in a worker (its GPU by default) and on the page's CPU and reports the differences ({maxDiff, meanDiff, over4, workerMs} in 8-bit levels): the smoke test uses it to check the GPU workers' pixels.

Disk cache

Settings ▸ Disk ▸ Disk Cache works in the browser, backed by the Origin Private File System (src/diskcache.rs): viewer frames are kept under anima-cache/v1/frames/<32 hex>.ecc between visits, in the desktop cache's format (disk_cache::frame_entry: 8-bit premultiplied RGBA, LZ4, a checksum), keyed by the same content hash (the comp and everything it uses, frame, scale, view, render options).

• A render request to a frame worker carries the frame's disk key; the worker posts the frame, then writes the entry with a synchronous access handle into a temporary file moved into place (cacheWrite in js/host.js; where move() is missing, the entry is written in place and its checksum catches a partial file), and reports FrameReply::Stored.
• The page keeps the index (DiskIndex, shared with the desktop's cache): rebuilt at startup from the folder listing (oldest files first, stale temporaries removed), updated by the workers' writes and by reads. The size limit is Maximum Disk Cache Size, at most half the origin's quota; the least recently used entries are deleted beyond it.
• A frame whose key is indexed is read and decoded on the page (asynchronously) instead of being rendered; a damaged or missing file is a miss and is forgotten. The timeline's blue cache bar shows the frames on disk.
• cache.diskStats reports the browser's cache ({enabled, available, loaded, entries, frames, layers, bytes, maxBytes, hits, misses, writes, evictions, layerHits, layerWrites}); Empty Disk Cache / edit.purge {what: "disk"} clears it. anima.info().diskCache has the same.

Layer buffers (M13.30). Slow layer buffers (≥ 20 ms to render) go to the disk cache too (anima-cache/v1/layers/<32 hex>.ecc, disk_cache::layer_entry, keys salted per project and footage as on the desktop). A layer lookup happens in the middle of a synchronous render, where the browser's file system can't be read, so it is made synchronous by fetching first:

1. Each frame worker backs its layer cache with a PrefetchStore (anima_render::cache): lookups are served from buffers fetched before the request; a lookup that finds nothing is recorded, and buffers to keep queue up as sealed entries.
2. The reply lists the frame's layer misses and the buffers served from the store (FrameReply::Frame layerMisses / layerHits); after posting the frame the worker writes the queued entries (sync access handles) and reports each (FrameReply::Stored layer).
3. The page keeps one index for frames and layers (DiskIndex: one least-recently-used order; hits touch entries) and remembers the workers' recent misses (disk_cache::LayerPrefetch). A render request names the recent misses that are on disk (FrameMsg::Render prefetch, at most 64); the worker reads those it lacks into its store before rendering, so their lookups hit (another worker's buffers, buffers from an earlier visit, buffers dropped by a purge).

Storage manager

Settings ▸ Disk ▸ Browser Storage (shown in the browser only) reports where the data lives (Origin Private File System, IndexedDB or memory), the origin's usage and quota (navigator.storage.estimate()), what the store holds (projects, imported media, auto-saves: count and size) and the disk cache, whether the storage is persistent (with Request Persistent Storage, navigator.storage.persist()), and Clear buttons for the disk cache, imported media, auto-saves and saved projects. Agents use the same through engine commands (crates/engine/src/storage.rs, the session's StorageHost):

The widgets have automation ids settings.storage.backend, .usage, .diskCache, .persist and .clear.<what>.

Offline and install (PWA)

manifest.webmanifest (name, icons, standalone display) makes the app installable; sw.js precaches every file of the build into a cache named after the build's hash, serves them cache-first (the cached responses keep the server's COOP/COEP headers, so the page stays cross-origin isolated offline), and drops older caches when a new build activates. After the first visit the app loads without a network.

window.anima: the agent / test API

The control channel of the desktop app (docs/control-protocol.md) as promises. Results resolve with the method's result, errors reject with the message.

window.animaLoad holds {wasmMs, readyMs} (module fetch + compile, and until the app runs).

await anima.execute("layer.newSolid", {color: "#ff8800"});
await anima.request("ui.menu.invoke", {id: "renderQueue.add", params: {format: "gif", resolution: 0.25}});
await anima.execute("renderQueue.render", {wait: false});  // renders in a worker; the GIF downloads
await anima.saveToBrowser("/my-project.ecproj");            // File ▸ Open Recent has it after a reload

Browser test

apps/anima-web/tests/smoke.mjs drives headless Chrome over the DevTools protocol (Node ≥ 22, no npm packages): load, the GPU path (WebGPU → GPU compositor, viewer on the GPU), the demo comp in the viewer, render.frame, Render Queue GIF and PNG sequence (.zip) downloads, a background render in a worker while measuring the page's event-loop gaps (must stay under 400 ms) and its progress, the AudioContext starting on a (simulated) key press and a tone playing through it with the meters moving, Save As and re-opening the saved project, a script opening a file from browser storage and the footage check finding stored media, persistence across a reload (session snapshot, a stored project, imported media, a setting, Open Recent) and an offline reload through the service worker. It writes screenshots and report.json:

cargo xtask web --serve 8765 &
node apps/anima-web/tests/smoke.mjs --url http://127.0.0.1:8765/ --out target/web/smoke

apps/anima-web/tests/workers.mjs checks the job-worker plumbing (js/host.js, web/worker.js) under Node with a fake Worker, without a build: a replaced file is sent to a reused worker again even at the same size, a worker that fails to start fails its job and is terminated, and a new Worker that throws (a sandboxed iframe) fails the job and reports the frame worker dead instead of stopping startup (node apps/anima-web/tests/workers.mjs).

apps/anima-web/tests/page.mjs checks page-side helpers of js/host.js the same way with a fake DOM: a picked file that can't be read is reported while the others are added, and the device-lost notice saves, flushes before reloading and reports a failed flush (node apps/anima-web/tests/page.mjs).

apps/anima-web/tests/storage.mjs checks the IndexedDB storage backend of js/host.js with a fake IndexedDB: a put or delete resolves only once its transaction has committed, and one whose transaction aborts after its request succeeded (or whose request fails or throws) rejects and leaves the stored entry as it was, so the flush keeps it for a retry (#1242) (node apps/anima-web/tests/storage.mjs).

It also checks the M13.10 paths: viewer frames rendered in frame workers while scrubbing with the CPU renderer (project synced as diffs, event-loop gaps under 400 ms), a wait: true render that replies when done while the page keeps running, and the Media Browser over browser storage; and the M13.24 ones: GPU effects (Gaussian Blur, Glow) rendered on the frame workers' own WebGPU devices in passes while the page's event loop stays free, the worker's GPU frame matching the page's CPU render (workerFrameCheck), the storage manager (storage.info, storage.persist, the Settings ▸ Disk widgets), and the disk cache: frames written to the Origin Private File System, served from it after a reload without rendering, and cleared with storage.clear; and the M13.30 ones: a Render Queue job and a Warp Stabilizer analysis on the job worker's own WebGPU device (workers.jobs: GPU passes and readbacks per job) while the page stays responsive, and layer buffers written to the disk cache and served from it after a reload and a purge (layerHits); M13.32: editing after the Warp Stabilizer analysis (removing it, pre-composing, a mask) keeps the page's event-loop gaps under 50 ms.

Native unit tests cover the storage model (apps/anima-web/src/store.rs: write coalescing, file table, ConfigStore / FileOps, auto-save and crash recovery through the browser store) and the worker protocol (crates/engine/src/offload.rs: serde round trips of every request and reply, a render through an in-process offload; tests_roto.rs: Roto Brush propagation through an offload streaming segmentations; tests_mask_warp.rs: Warp Stabilizer plans solved in the job, never by warp.status on the page), the frame-worker protocol (crates/engine/src/remote.rs: diff / patch, a replica following edits renders the same pixels; crates/ui-egui/tests/it/ui_web_offload.rs: the frame cache through a remote renderer, wait: true deferred until the offloaded job ends) and the Media Browser's virtual tree (media_browser.rs, tests_panels.rs). M13.24: deferred readbacks on a real device (crates/gpu/src/tests_deferred.rs: whole-frame GPU and CPU-composited frames in passes match the CPU, placeholders never reach the layer cache; deferred.rs: keys, late readbacks of an older frame dropped), the frame server's passes and Auto timing (remote.rs, a deferred stand-in accelerator), the browser's routing and disk cache (crates/ui-egui/tests/it/ui_web_gpu_storage.rs: a GPU frame worker renders effect frames, the page's GPU keeps frames without effects, frames stored under their disk keys come back after a "reload" without rendering, the blue cache bar sees them; the storage manager's widgets and commands), the disk index (disk_cache.rs), the store's usage by kind and clearing (store.rs), storage.* (storage.rs), and every WGSL float literal fitting f32 (browsers reject the whole module otherwise; tests.rs). M13.30: job workers on a deferred device (crates/ui-egui/tests/it/ui_web_job_gpu.rs: a Render Queue job and a Warp Stabilizer analysis render their frames in passes on the device, with the CPU worker's pixels / result), particles, Advanced 3D and the job pass driver in passes (crates/gpu/src/tests_deferred_jobs.rs), layer buffers through the disk store and the prefetch plan (remote.rs layer_buffers_go_through_the_disk_store).

Gaps

• No shared-memory threads: each worker is a separate engine instance (memory per worker: the project replica and the footage it reads); see the threading design above.
• Browsers without WebGPU in workers render frames, Render Queue jobs and analyses on the workers' CPU. A frame renders at most 24 passes on the GPU, then on the CPU.
• A layer buffer is fetched from disk only after some worker missed it in a recent frame (the first frame after a reload renders its layers).
• Cancelling an analysis drops its partial result (the worker is terminated).