Turn talking-head footage into a designed film with a coding agent. Recording + complete SRT + brief → MP4 + editable Three.js project.
Demos · Get started · How it works · Compatibility · Membership & help
English · 简体中文 · 日本語 · Español · Français
Greenery, explained · 60 seconds · Original left / result right
Three owner-approved films with matching source timecodes, tracked soft-edged face blur and silent previews. Versions, excerpts and viewing scope
1. Download and install
You only need Node.js 22+ and a coding agent. Download this repository with Code → Download ZIP, then extract it; Git is not required. Open the extracted folder and double-click start.command on macOS or start.bat on Windows. On Linux, or from your agent's terminal, run:
node scripts/bootstrap.mjsnpm run onboard runs the same installer. It prepares the project dependencies, local Python environment, Chromium, FFmpeg / ffprobe and the person-matting model, then checks the environment and runs a short matting test. The first run needs an internet connection; let it finish. You do not need to install Python separately or tune a model. Step-by-step guide and troubleshooting
2. Give your agent the brief
Put your recording and complete SRT in inputs/, open the extracted folder in your coding agent, and send:
Read AGENTS.md, follow the installation guide if needed, then read the two project-local Skills. Create a complete talking-head film using inputs/source.mp4 and inputs/source.srt. Prepare and verify a synchronized person cutout before composition; keep the original speech order and timing. Work in work/my-first-film/. Complete every production stage and inspect the actual picture and audio. Deliver the MP4 and editable project. Respond in English.
Person matting is required for talking-head films. The installer supplies a local CPU-capable method; a suitable existing person layer or another compatible tool can replace that method. A dedicated GPU is not required. Image generation remains optional.
3. View the result
Ask your agent to open the finished MP4 and editable preview. It can run npm run serve and open the film page it created at work/my-first-film/. The server root has no preset film.
Checks your agent can run
npm run doctor -- --matting
npm test
npm run smokeDoctor checks the matting environment; the rendering smoke performs a real Three.js → H.264 render. Neither replaces reviewing the person edges and the finished film. Installation and these local checks do not call paid generation services.
Want project resources, video examples, or written Q&A with the author? Erduo’s SURGE (无限涌动) offers Seed Membership with lifetime membership status, project resources, video examples, ongoing updates, and community text Q&A. You can also connect your own AI to the member library within your access rights. Available resources, benefits, and prices are listed on the website.
Explore membership · Sign up / sign in and purchase · Discuss custom services
Member text Q&A excludes remote installation, hands-on engineering troubleshooting, and one-to-one production work. Video production, deployment, training, and special licensing require a separate scope and quote. Membership does not replace this repository’s license or automatically grant commercial rights.
Contact the author through the website. Existing members can open the member library.
Analyze the complete subtitles, then prepare assets and the person layer, compose type and space, animate continuously, mix sound, and review the output.
One continuous author develops the opening and hardest passage before expanding the film. The agent reads each stage, preserves speech timing and inspects the actual picture and sound.
Design principles: meaning, people, typography and continuous motion
- Meaning first: choose actions that explain the spoken idea; keep evidence and timing intact.
- A person in space: keep a synchronized person layer, visible depth and intentional framing. Outline only when replacing the original background.
- Typography with a role: compare fonts in real compositions; establish hierarchy, readable landings and reading time.
- Continuous movement: carry objects and attention between ideas. Validate a continuous sample before expanding the film.
Agents: Codex, Claude Code, Gemini CLI, Cursor and GitHub Copilot have entry instructions for the same local Skills. Other agents can read AGENTS.md directly. No Codex-specific API is required. Hosts need file, terminal, browser and audiovisual inspection capabilities.
Platforms: all six Windows, macOS and Ubuntu × Node 22/24 CI jobs passed automatic setup, real CPU matting and video-rendering checks on September 20, 2026. Full-film production is evidenced in Codex; the other clients have not each completed an independent full-film test. Compatibility and test evidence
- Architecture — the two Skills, toolchain and repository structure.
- Reference analysis library — 43 cases, 109 mechanism windows and five tutorial methods; text / JSON only, without reference footage, audio, frames or complete transcripts.
- Contributing — how to make and verify changes.
READMEs are available in five languages; canonical Skills are maintained in Chinese. Agents can work in your language. Check fonts, glyphs, line breaks and reading time for each film; automatic dubbing is not included.
Noncommercial use. Commercial use requires prior written permission.
Project-owned code, Skills, documents and analysis use Motion Craft Community License 1.0, a custom license based on Apache-2.0 terms with noncommercial restrictions. Fonts and dependencies retain their licenses; demo footage is excluded from the project grant.