ai-video-editor

# Timeline Studio — Full Product Reference

Last verified: 2026-09-10
Canonical URL: https://video-editor.ai-creator.top/
GitHub: https://github.com/MartinDelophy/ai-video-editor
License: MIT

## Definition

Timeline Studio is a free, open-source, local-first AI video editor that runs in a desktop web browser. It opens directly into a multi-track editing workspace for importing local media, generating narration and captions, editing visual clips, previewing a composition and exporting video.

## AI voiceover

- Chinese and mixed Chinese/English: Hojo TTS Light 80M FP16 WebGPU with two built-in female reference voices (晴岚, 若溪).
- English: Kokoro 82M v1.0 ONNX q8 on WASM.
- Additional browser Piper voices: German, Spanish, French, Italian and Brazilian Portuguese.
- Generated speech is inserted at the current playhead as a voiceover clip.
- Browser microphone recording is also available after user permission.
- Authorized Chinese and English target-voice conversion uses OpenVoice V2 as a second stage after base synthesis.
- Saved clone profiles and their embeddings remain browser-local; completed conversions go to My assets and never replace timeline audio automatically.

## Automatic captions

- Default recognition model: Whisper small q8 ONNX.
- Execution: WASM worker for subtitle stability.
- Timing: coarse recognition boundaries are adjusted toward nearby source-audio energy.
- Chinese cleanup: conservative contextual correction for high-confidence homophone mistakes.
- Caption text and segments are edited in the right-side caption context panel.
- Global caption placement and styling remain in the caption tool.
- Preview and export share scale-aware caption layout.

## Visual editing

- Visual clips support scale, horizontal position, vertical position, rotation and opacity keyframes.
- Keyframes can capture all transform properties together or key individual properties independently.
- Supported masks: rectangle, rounded rectangle and circle, with position, size, feathering and inversion controls where applicable.
- Masks stay in canvas coordinates while the visual media transforms behind them.
- Video speed is clip-scoped and uses the same source-time mapping for preview, timeline duration, keyframes and export.
- Extracted source audio can stay linked to its video or use independent timing.

## Timeline

- Visual, sticker, caption, source-audio, voiceover and music tracks.
- Clip operations include selection, moving, splitting, duplication, resizing and deletion.
- Track operations include visibility and locking.
- Multiple generated voiceover or caption lanes are created when clips overlap.
- Timeline geometry remains time-accurate across zoom levels.

## Browser vision

- YOLOS tiny q8 ONNX detects subjects.
- MODNet q8 ONNX provides portrait matting.
- Subject geometry supports smart crop and caption avoidance.
- Video analysis produces timestamped samples across the clip instead of reusing only the first frame.
- Model inference is precomputed and bounded rather than performed inside the real-time export loop.

## Talking portrait

- JoyVASA ONNX converts audio into motion coefficients.
- LivePortrait performs neural rendering on WebGPU.
- 256px mixed-FP16 mode is intended for faster preview.
- 512px mixed-FP16 mode is intended for quality output and can be slow on current browser GPUs.
- Sparse neural keyframes are interpolated for encoded output.

## AI music and repair

- Browser-local AI music uses a prompt-first WebGPU workflow with translated descriptions, compact musical controls and persistent model caching.
- Completed AI music goes to My assets and is not inserted into the timeline automatically.
- AI Repair includes browser-local object or watermark removal, video restoration and denoising workflows with reversible asset application.
- Subject and depth analysis support editable person or object effects, smart framing, cinematic depth and photo parallax.

## Generation connectors

- Puter uses its real provider-managed browser session and user-pays generation flow.
- ComfyUI and Stable Diffusion WebUI are separate loopback-only connectors with real job or API output retrieval.
- Completed connector media goes to My assets and is not inserted into the timeline automatically.

## Export

- Export controls include resolution, frame rate, codec and quality.
- Available resolution choices include 720p, 1080p, 1440p and 2160p.
- Available frame-rate choices include 30fps and 60fps.
- MP4 is produced when the browser supports the required native or browser-side conversion path.
- WebM is used as a fallback when MP4 conversion is unavailable or fails.
- Export renders the project canvas and mixes active audio tracks in the browser.

## Privacy and networking

- Images, video and audio selected for a project are processed in the browser.
- Model and runtime files can be downloaded from configured third-party hosts.
- A service worker caches supported application and model assets for repeat use.
- Interface language and other preferences may be stored in browser local storage.
- Timeline Studio should be described as local-first, not as making zero network requests or bundling every model offline.

## Supported interface languages

Chinese, English, Japanese, Korean, Spanish, French, German, Portuguese, Thai, Vietnamese, Russian, Italian and Indonesian.

## Agent integration

- Native WebMCP is a progressive enhancement for compatible browsers and agent hosts. The normal editor remains usable without it.
- The 21 browser tools inspect the open project, tracks, clips, serialized captions, timeline markers and available editor assets; seek the preview; undo the latest unchanged agent edit; and download an editable .timeline copy.
- The 26 reviewed operations support caption add/update/delete/style/center-point position, project aspect ratio/fit, timed-clip move/resize/source-trim/delete, existing overlay transforms, audio/music volume and fades, muting, markers, main-visual split/delete/duplicate/reorder/trim, insertion of ready editor assets, and picture-in-picture insertion with an initial transform.
- Edit previews use an opaque live-state token and either an operation list or a legacy complete main-visual clip permutation. Apply requires the returned preview ID and rejects intervening project or asset changes. Split offsets are clip-relative; source trims use absolute original-source seconds. Complex timing, transitions, keyframes and effects constrain supported edits.
- Changes reuse the shared command engine, current ripple mode, track locks and editor history. Main visuals remain gapless, caption/audio associations are preserved, and marker range ends do not extend rendered content duration.
- Video export has prepare/start/inspect/cancel tools over the real editor pipeline. Prepare resolves settings, range and estimates; start revalidates the project and uses an idempotent request key. Inspect reports progress, actual extension, bytes, pipeline and any WebM fallback for an MP4 request. Cancellation reaches the encoder/transcoder. Browser download initiation is reported separately from verified disk persistence.
- Explicit media sampling returns up to four JPEG frames and/or a ten-second 24 kHz mono WAV mix from the shared composition renderer, with audio peak/RMS/clipping measurements and no playhead movement. Metadata tools continue to return no media bytes.
- Browser-local AI uses capabilities/prepare/start/inspect/cancel tools with current state checks, unique retry keys and explicit model-download permission. Voiceover generates up to 2,000 characters in stable-speaker breath groups and commits to My assets; transcription reads up to 120 seconds of a ready audio asset and proposes caption operations. Neither inserts results automatically. Transcription does not infer timeline speed-curve mapping, and cloned-voice enrollment/conversion is not exposed.
- Project and video delivery use browser downloads. Asset insertion references media already available in the editor; browser tools expose no arbitrary URL or filesystem import, media upload, arbitrary model execution, remote generation or cloud rendering.
- The separate repository Skill, CLI and local STDIO MCP operate on project files. The website does not host a remote MCP or OAuth server.
- Tool results can include project metadata, caption text and explicitly requested rendered media samples, which the browser agent receives under its own data-handling terms. Media bytes are not returned by inspection tools.
- Browser guide: https://video-editor.ai-creator.top/agent-guide.md
- Agent Skills index: https://video-editor.ai-creator.top/.well-known/agent-skills/index.json
- Browser Skill: https://video-editor.ai-creator.top/.well-known/agent-skills/edit-timeline-studio-browser/SKILL.md
- Full local Skill: https://github.com/MartinDelophy/ai-video-editor/tree/main/skills/edit-timeline-studio
- Integration reference: https://github.com/MartinDelophy/ai-video-editor/blob/main/docs/webmcp.md

## Technology

React 19, Vite 6, ONNX Runtime Web, WebAssembly, WebGPU, MediaRecorder, browser Canvas APIs and FFmpeg WASM support.

## Official factual pages

- https://video-editor.ai-creator.top/features/
- https://video-editor.ai-creator.top/how-it-works/
- https://video-editor.ai-creator.top/faq/
- https://video-editor.ai-creator.top/privacy/
- https://video-editor.ai-creator.top/about/
- https://video-editor.ai-creator.top/ai-video-editor/
- https://video-editor.ai-creator.top/ai-voiceover-generator/
- https://video-editor.ai-creator.top/automatic-caption-generator/
- https://video-editor.ai-creator.top/browser-ai-music-generator/
- https://video-editor.ai-creator.top/local-ai-video-editor/
- https://video-editor.ai-creator.top/talking-avatar-generator/
- https://video-editor.ai-creator.top/articles/