# Ghost Director

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Muratcan Koylan](https://cerebralvalley.ai/u/koylanai), [Ata Koylan](https://cerebralvalley.ai/u/atakoylan)
- **GitHub:** https://github.com/muratcankoylan/ghost-director
- **Demo video:** https://youtu.be/xQW2XQsdw0c
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3276

Ghost Director is an open-source directing workspace that gives video models a real 3D scene to follow.

The problem: AI video starts each shot from a text prompt and a few images, with no persistent model of the room. "Slow push-in toward the pitcher" leaves the start pose, end pose, focal length, speed and the actor's position to the model, and the next shot re-guesses all of it. In commercial work, where a client has approved a character, a product and a location, that drift means rewriting prompts and paying for more takes.

How it works: Ghost Director keeps one explicit scene for the whole production.
1. Muse Image generates reference sheets for characters, props and locations. Approved versions are content-hashed and pinned.
2. SAM 3D reconstructs textured meshes from those approved images. The director sets real height and orientation, Blender normalizes each mesh to meters, and each mesh is compared against its photos before it can be placed in the set.
3. Camera and blocking are one timeline in meters and seconds, shared by the Three.js preview and Blender. Blender renders it locally as a control video.
4. The video model receives the approved photos for appearance and the control video for motion, with a map binding each proxy in the video to its references. The director approves the exact request and price, bound to a SHA-256 hash.
5. Muse Wonder roles plan shots and check takes against continuity assertions. Accepting a take writes its end state back as a new scene revision for the next shot.

What works: on our test scene, three references and two meshes cost about $0.07 to generate, a 12-second control video rendered locally in 11 seconds, and Blender matched the plan within 1.2e-7 at sampled timestamps. An earlier product test produced a generated clip from a Muse Image reference and a Blender camera move. 

Team: Ata Koylan is a generative AI filmmaker who makes AI commercials and independent film; he directs. Muratcan Koylan is a Member of Technical Staff at Sully.ai and author of Agent Skills for Context Engineering (18,000+ GitHub stars); he builds the system.

## More from Meta Global AI Developer Hackathon

- [Prashanth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3273)
- [Oc2cO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3274)
- [Chidimma Joel](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3275)
- [NPC Worldwide](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3277)
- [Zambrite](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3278)
- [crosscheck](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3279)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3276. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt