Delivery

Quick export offers Draft, Share and Master presets. Draft selects up to 720 px and a fast low-quality H.264 encode; Share selects up to 1080 px at high quality; Master keeps the authored canvas at maximum quality. Presets never upscale a smaller document. Deliver opens the encoder controls by default.

Encoder controls

MP4 supports H.264 and H.265, an explicit CRF from 0 to 51 or a positive target bitrate, a closed set of encoder speed presets, and yuv420p, yuv420p10le or yuv444p output. CRF and bitrate are mutually exclusive. The default remains H.264, quality high, medium preset and yuv420p; absent knobs preserve the existing argument list and serialization.

const settings = Export.mp4(
  codec: ExportCodec.h265,
  crf: 21,
  preset: EncoderPreset.fast,
  pixelFormat: ExportPixelFormat.yuv420p10le,
);

The dialog’s choices affect one render snapshot. The saved document changes only when explicitly edited. Frame-rate changes are document edits because all timing is measured in composition frames.

Desktop and browser FFmpeg receive typed, validated argument arrays. Browser codec availability depends on the installed FFmpeg build; an unavailable codec fails visibly. Mobile hardware export accepts codec and target bitrate and rejects software-only CRF, preset and pixel-format requests before capture.

Cancellation and cleanup

Each export can carry a RenderCancellation token. Desktop capture checks it between frames; encoding terminates the FFmpeg child process and waits for it to exit before deleting the partial sandbox. Failed renders also remove their sandbox. Successful desktop delivery copies the completed artifact to the chosen location and then removes the temporary render files.

Browser cancellation terminates the FFmpeg worker, resets its loaded state and clears the in-memory capture sandbox. A later job reloads the worker. Successful and failed encodes delete staged virtual input/output files. The browser bridge must expose terminate and deleteFile for cancellable jobs; unsupported bridges fail explicitly. Disposing the desktop capture surface unmounts the widget subtree before releasing the render pipeline.

Maintained tests exercise typed argument combinations, legacy defaults, round-trip digests, actual H.264/H.265 encodes, real process cancellation, partial-sandbox removal, browser worker reload/cleanup, and desktop capture subtree disposal.

The maintained browser acceptance runs the app’s real local FFmpeg worker and WebCodecs bridge, including encode, cancellation, reload, staged-file cleanup, first/middle/last source-frame pixels, embedded-audio metadata, and bounded RGBA retention. Run dart apps/slides/tool/verify_web_encoder.dart after vendoring browser assets. Source video preview currently demuxes MP4/MOV ISO BMFF files; other containers report that conversion is needed. Browser codec availability also depends on WebCodecs support for the file’s video codec.

Preview quality and background exports

Full, Half and Quarter bound the preview’s decoded clip raster. A 4K RGBA frame uses about 31.6 MiB at Full, 7.9 MiB at Half and 2.0 MiB at Quarter. Scrubbing keeps a bounded recent-frame cache. Layout and source timing stay the same. Draft can bypass effects and grades; the stage explicitly labels that choice. These preferences never modify the project or an export snapshot.

Add exports to the Deliver queue or choose Batch 1080p + 720p. Each job owns the project snapshot taken when queued, so later edits do not change an ongoing render. Jobs run in order, display progress and can be cancelled individually. A failed job leaves following jobs runnable. The queue survives workspace switches and belongs to the current editor session, rather than disk storage.

The repeatable optimized performance checks are bash tool/verify_editor_performance.sh profile and bash tool/verify_editor_performance.sh release. They drive a three-minute 4K project with eight lanes and 240 placements sharing one six-second source on Linux, measuring loading, digest work, timeline gestures, canvas drag, source decoding and preview capture. Engine frame timings are recorded separately from test-driver latency. Reproduction steps and measurement limits are in the release verification guide. The complete native journey saves and reopens identical project data, continues editing during rendering, then verifies both real MP4 artifacts with ffprobe.