That is convenient for the application developer, but it changes the product for the user. Raw footage can be large, personal, slow to transfer, and expensive to process repeatedly. For a video editor, uploading is not a one-time action either: every trim, caption, effect, and export can create another round trip.
I wanted to explore a different constraint: how much of an AI video-editing workflow can stay inside a modern browser?
That question became Timeline Studio, an open-source, local-first AI video editor with a multi-track timeline, captions, voiceovers, visual effects, browser AI tools, and offline export.
The interesting part was not putting a familiar editor UI on a webpage. It was making media, AI inference, model delivery, timeline state, preview, and export behave like one coherent system.
For supported workflows, project media is processed in the browser instead of being uploaded to an editing backend. This changes several engineering decisions at once: inference needs browser-compatible runtimes; model downloads must be lazy, resumable, cached, and versioned; timeline operations need deterministic state rather than server-side session magic; preview must stay responsive while export remains reproducible; failures need to explain browser or hardware limitations without hiding them.
WebGPU is the main compute path for heavier models. WASM remains important for operations where it is more stable or broadly compatible. ONNX gives the project a practical interchange format, but converting a model is only the beginning. A model that technically loads can still be unusable because of memory pressure, unsupported operators, slow graph initialization, or inconsistent numerical output across execution providers.
The result is a mixed runtime rather than a single “run everything on WebGPU” switch.
Timeline Studio lazy-loads AI artifacts only when a feature is first used. The service worker owns the persistent model cache, while later runs reuse the cached files. Models are pinned to known revisions so an upstream change cannot silently alter an editing workflow.
The project also mirrors model artifacts across Hugging Face and ModelScope. Chinese and domestic sessions prefer ModelScope; other sessions prefer Hugging Face. If the preferred provider is unavailable, the application can fall back to the other mirror.
The important detail is that both providers map to one cache identity. A fallback should not create a second full copy of the same model on the user's device.
This sounds like infrastructure work—and it is—but it directly affects the editing experience. “Generate music” should not look like a fresh model download every time. A cache miss should not become a mysterious network exception. A pinned model should carry its license and source notes with it.
Video editors need immediate feedback. Native media playback and a live canvas make trimming and positioning feel responsive, but they are not enough for final export.
Real-time playback can vary with decoding speed, dropped frames, tab scheduling, and device load. If final rendering simply records whatever happened on screen, two exports of the same project can drift.
Timeline Studio therefore treats the timeline as declarative project state and uses a separate offline rendering path for export. Clips, source trims, transforms, captions, overlays, audio, effects, and keyframes are evaluated against explicit timestamps. WebCodecs is used for MP4/WebM composition where available, with a compatibility path for browsers that need it.
The two paths have different performance goals: Preview: respond quickly enough to edit. Export: evaluate the same project state consistently.
They still need shared geometry, timing, interpolation, and effect definitions. Otherwise the editor develops the most damaging kind of bug: a preview that looks correct and an export that does not.
This is why small visual features can require surprisingly deep work. A timed ripple effect, for example, is not just a CSS animation. Its seeded placement, beat timing, grayscale-to-color reveal, wave propagation, and decay must all be derived from the same deterministic parameters in both preview and export.
Another design decision was to avoid treating AI as a collection of isolated demo buttons.