A browser video editor pays for its missing upload button in runtimes, caches and determinism

The reason this is not a skin-deep change is arithmetic on round trips. In an editor, an upload is never a single event: according to the project's writeup, every trim, caption, effect and export can create another transfer ^[15]. Take the media local and those transfers stop, but the work does not disappear. It moves into the tab, where the developer now owns inference, model delivery, caching and determinism. The writeup names five commitments that follow from that one decision ^[7][17].

The runtime story is the least glamorous and probably the most instructive. WebGPU is the main compute path for heavier models, with WASM kept for operations where it is more stable or more broadly compatible, so the result is a mixed runtime rather than one switch ^[3]. ONNX gives a practical interchange format, but the author is blunt that conversion is only the start: a model can load and still be unusable through memory pressure, unsupported operators, slow graph initialisation, or numbers that differ across execution providers ^[4]. That last item is the expensive one, because it turns model adoption into per-provider qualification rather than a packaging step.

Delivery is where local-first quietly becomes a storage contract with the user. Artifacts load lazily on first use of a feature, a service worker owns the persistent cache, and later runs reuse the files ^[8]. Revisions are pinned so an upstream change cannot silently alter an editing workflow ^[9], and artifacts are mirrored across Hugging Face and ModelScope, with Chinese and domestic sessions preferring ModelScope and others preferring Hugging Face, plus fallback between them ^[10]. The detail worth stealing is that both providers resolve to one cache identity ^[18]: two mirrors, one copy on disk, one avoided duplicate per model ^[19]. Availability routing that would normally be a CDN concern here lands on someone's laptop.

Then determinism. Real-time playback varies with decoding speed, dropped frames, tab scheduling and device load, so an export that records whatever happened on screen can drift between runs of the same project ^[12]. The answer is a declarative timeline evaluated against explicit timestamps through a separate offline render path, with WebCodecs for MP4 and WebM composition where available and a compatibility path elsewhere ^[13][14]. Two paths, one evaluation model: preview and export must share geometry, timing, interpolation and effect definitions, or the editor grows the bug where the preview is right and the file is wrong ^[5]. Hence the claim that a timed ripple effect needs seeded placement, beat timing, the grayscale-to-color reveal, wave propagation and decay derived from the same parameters on both paths ^[6]. Small feature, shared kernel.

One source, no benchmarks, no model sizes and no export timings, so treat the performance claims as design intent rather than measurement. The falsifiable part is structural: whether preview and export genuinely call the same evaluation code, and whether AI output really returns to the editing model instead of sitting behind demo buttons ^[16].