ai-video-editor

Timeline Studio — Browser AI Video Editor

English | 中文 | 日本語 | 한국어 | Español | Français | Deutsch | Português | ไทย | Tiếng Việt | Русский

Create with Anna

Try the Anna edition for AI-assisted editing and project continuity:

• Describe your edit to Anna, review a plan based on your brief and clip metadata, then apply supported clip ordering and simple 1× video trims.
• Projects and media save automatically to Anna cloud storage and can resume with the same Anna account in another browser. Local project checkpoints provide an additional recovery option.
• Refine the editable multi-track timeline, add AI voiceovers and manual captions, and export MP4 using an available browser encoder.

This desktop preview plans from metadata, without analyzing video frames or audio; local AI features depend on the Anna container’s capabilities.

→ Try Timeline Studio on Anna

Responsible use of deep synthesis

This tool uses deep-synthesis technology and is intended solely for technical research and learning.

Users must ensure that they:

• use only facial images or videos of themselves or people who have provided lawful authorization;
• do not create or distribute any illegal, infringing, false, or misleading content;
• do not present generated content as authentic footage or impersonate another person without their consent.

Users are solely responsible for any legal liability arising from violations of these requirements.

Project updates

• 2026-10-13 — Agent relighting and depth-video conversion now reuse the production browser pipeline through WebMCP prepare/start/inspect/cancel. Lighting parameters use shared reviewed, undoable commands; generated depth videos stay in My assets until a separate reviewed insertion. The editing Skill documents the complete workflow and local CLI limits.
• 2026-10-12 — Long-video playback pauses background filmstrip refinement and draws the exact playhead thumbnail without JPEG encoding. Resume prioritizes the latest seek; AAC export preparation copies audio packets without re-encoding or copying the full video into WASM. Visual effects and export settings are preserved. Export now prefers supported hardware encoders and feeds eligible unscaled video frames directly to the compositor, preserving the existing raster path for transforms and effects. Audio encoding and next-frame composition now overlap video submission with a bounded queue. Scrubbing separates decoder/playhead updates from editor reconciliation, preserves marker snapping and commits the exact release position.
• 2026-10-07 — Visual clips now support cuts down to 0.05 seconds, fixing rejected cuts at valid short intro positions such as 0.45 seconds. Trimming, speed changes and project recovery share the same minimum duration. Fixed frozen video playback when entering or leaving the large canvas preview by restoring source timing, playback and frame callbacks on the new video element.
• 2026-10-05 — Depth sampling now uses Fast 8, Standard 16 and Fine 24 samples/s. Fast and Standard use RIFE to output 24 fps while sampling and encoding overlap in a bounded pipeline. Source audio is preserved; completed analysis remains reusable after generation failure. Fixed sampling clocks and reuse of RIFE sessions, pair pixels and tensors reduce redundant interpolation without lowering precision, resolution or sample counts. RIFE now downloads from owned, immutable Hugging Face and ModelScope mirrors with verified checksums, automatic fallback and a shared cache.
• 2026-10-04 — Depth map — Turn video into grayscale depth: near is bright, far is dark. Depth analysis overlaps frame decoding with inference, encodes PNGs in the worker, and throttles UI updates. Depth-map preview and export now interpolate between samples, match trimmed and retimed source timestamps, and reuse decoded images. Depth-map video saved to My assets. Generated depth videos replace their source clips while retaining assets; playback duration follows the visual sequence, fixing zero duration after clearing and reinserting video. Depth videos now include source audio matched to trimming and speed changes; the original video and its sound remain available. Editable depth-assisted relighting adds a lamp-position sphere and hold-to-compare preview, with refreshed frames after timeline dragging. See the public Roadmap for planned work, Releases for shipped changes, and Issues for focused tasks and bugs.

What can it produce?

Explore reproducible before/after examples and editing recipes:

→ AI Video Editing Skills Handbook

Timeline Studio is a local-first AI video editor that runs in the browser. It combines a CapCut-style multi-track timeline with WebGPU AI music and repair, multilingual voiceovers, automatic captions, talking-avatar generation, and deterministic offline export.

Open the editor · Watch on YouTube · Hugging Face Space

Video demo

Auto Edit

Visual analysis, keyframes, captions, and export.

https://github.com/user-attachments/assets/e8327caa-429e-40ff-a7fe-a59e6cf7a464

AI Repair

Timed watermark removal with before/after review.

https://github.com/user-attachments/assets/aea9f5b4-c720-4b0c-9067-5ec124eef982

AI Voiceover

https://github.com/user-attachments/assets/304a744e-d620-4380-9c17-19af3726f5a4

AI capabilities

• Multilingual voiceover: Chinese and mixed Chinese/English Hojo TTS Light 80M FP16 reference voices (晴岚 and 若溪), with autoregressive generation on WebGPU and stable waveform decoding on WASM; English Kokoro 82M; and browser Piper voices for German, Spanish, French, Italian, and Brazilian Portuguese.
• Local AI music: Stable Audio 3 Small Q4 ONNX runs through WebGPU with translated free-form prompts, 30/60/90/120-second choices, waveform-aware long-track looping, persistent model caching, and automatic insertion into My Assets.
• Automatic captions: Whisper small q8 ONNX with waveform-aware timing and conservative Chinese recognition cleanup.
• Smart framing: YOLOS tiny subject detection and MODNet portrait matting for smart crop, caption avoidance, and background removal across images and complete videos.
• AI Repair: browser-local MI-GAN watermark/object removal with multiple timed repair regions, plus NanoVSR 644K WebGPU 4× restoration for images and videos with synchronized before/after comparison.
• AI vocal separation: isolate vocals and place the instrumental stem on the music track without leaving the browser workflow.
• Digital human: JoyVASA audio-to-motion and LivePortrait neural rendering with WebGPU, 256px preview and 512px quality paths.
• Local-first inference: large models are lazy-loaded, revision-pinned, and cached by the service worker; supported workflows run without uploading project media to an editing backend.

Resilient model delivery

Browser AI models are mirrored on both Hugging Face and ModelScope. Timeline Studio uses ModelScope first for Chinese interfaces and domestic sessions, uses Hugging Face first elsewhere, remembers the first working source for the current runtime, and automatically falls back to the other source when needed. Cache identities are shared across both providers, so switching routes does not download the same revision twice.

• Stable Audio: Hugging Face · ModelScope
• Voice models: Hugging Face · ModelScope (mixed upstream licenses; see the repository model card)
• Timeline Studio ONNX models: Hugging Face · ModelScope
• Depth Anything V2 Small Q4F16: Hugging Face · ModelScope (Apache-2.0)
• Vocal remover: Hugging Face · ModelScope

Editing and export

• Contiguous main Visuals track plus timed picture-in-picture overlays.
• Direct canvas selection, movement, proportional resize, rotation, masks, filters, effects, animation, speed, and explicit keyframes, plus desktop four-way color grading with fully keyframeable temperature, tint, saturation, hue, wheel saturation, and luminance controls.
• Captions, stickers, voiceover, separated source audio, and music on independent timed tracks.
• CapCut-style snapping, alignment guides, clip menus, split/duplicate/delete, timeline zoom, undo/redo, and portable .timeline projects.
• Native media playback for a responsive preview; export uses a separate deterministic offline rendering path.
• WebCodecs MP4/WebM composition with shared preview/export geometry, audio mixing, captions, overlays, effects, and MediaRecorder fallback.
• Installable PWA with a cached app shell and multilingual UI.

Agent Skill

The repository includes the AI Video Editing Skill for Codex, Claude Code, Copilot and Gemini CLI, backed by edit-timeline-studio for planning, executing, and verifying editable video timelines.