I’m building Timeline Studio, an open-source, local-first video editor that runs in the browser. The interesting part of this project is not the number of features. It is whether the browser can provide a reliable foundation for serious media workflows.
One of the biggest questions is deterministic export. Given the same project, media files, and settings, should the editor always produce the same frames and timing? In practice, browser APIs, hardware encoders, GPU drivers, and codec implementations can behave differently. A useful definition of “deterministic” may therefore need to distinguish exact binary output from visually and temporally identical output.
Preview and export consistency is another challenge. WebCodecs makes low-level decoding and encoding possible, but interactive preview and offline export often follow different execution paths. Frame selection, timestamps, speed changes, transitions, color processing, and audio synchronization must use the same timeline logic. Otherwise, the exported video can differ subtly from what the user approved in the preview.
WebGPU and ONNX introduce another layer of uncertainty. Performance may vary significantly between NVIDIA, AMD, and Intel GPUs, as well as between Windows, macOS, and Linux. A model that works well on one system may encounter unsupported operations, memory pressure, driver issues, or slower fallback execution on another. I’m especially interested in real-world results from Linux and integrated-GPU machines.
Large browser-local models also create product problems beyond inference speed. The first download can be several gigabytes, cache capacity is not always predictable, and GPU memory can be exhausted when multiple models or sessions remain active. The editor needs understandable download progress, reusable caching, controlled session lifetimes, and graceful recovery when storage or GPU memory is insufficient.
Keeping media on the user’s device provides a meaningful privacy advantage: footage does not need to be uploaded to an editing server. But “local-first” is not the same as “automatically private.” Remote model providers, analytics, browser extensions, external assets, and cloud-connected features can still cross that boundary. The interface should make those boundaries explicit instead of presenting local processing as an absolute guarantee.
I’m also experimenting with portable .timeline
project files and reversible Agent editing. The goal is for an Agent to inspect a project, propose semantic changes, and apply them without flattening or destructively replacing the original edit. Agent actions should remain understandable, reviewable, and reversible—closer to structured editing commands than automated video generation.
Finally, there are honest platform limitations. The project currently depends heavily on Chromium because WebCodecs and WebGPU support is more complete there. Firefox and WebKit support is still insufficient for parts of the workflow, and behavior may differ even among Chromium-based browsers. This is a real constraint, not something a compatibility badge can hide.
If you have experience with WebCodecs, WebGPU, ONNX Runtime Web, browser storage, GPU compatibility, or deterministic media pipelines, I’d value your feedback—especially test results from less common hardware and Linux setups.
GitHub: [MartinDelophy/ai-video-editor](https://github.com/MartinDelophy/ai-video-editor)