# NeoHuman

- **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:** [Serge Hendrickx](https://cerebralvalley.ai/u/SergeHendrickx)
- **GitHub:** https://github.com/SergeHendrickx/NeoHuman.git
- **Demo video:** https://drive.google.com/file/d/1j-Ypx-zqA2n8pHu-xrXhvrf-qoKPnCX4/view?usp=drive_link
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5183

NeoHuman — one photo, one parameter vector: a character you can see and hear
The gap
Single-photo human reconstruction has become impressive and almost useless in the same breath. A model returns a mesh — a beautiful, frozen shell with no controls. You cannot open its jaw, change its expression, pose it, or hand it to an animator; you can only look at it. And even when the mesh is riggable, the character still has no voice. Visual tools build a head, audio tools pick a preset from a list, and nothing connects them.

NeoHuman takes a different output seriously: not a mesh, but the parameters that produced it. From one photograph we recover a metric, animation-ready character on Meta's MHR parametric body/head model — 45 identity, 130 pose, 28 scale and 72 expression coefficients — and we treat that coefficient vector as the character's interface. The same 72 expression numbers that open the mesh's jaw are the articulation input for a speech decoder, so a jaw edit is simultaneously a geometry edit and a sound edit. The visual pipeline that recovers those numbers is what we ship in this repository; the voice side consumes the vector in the demo.

Supervision: Sapiens2
Everything starts with one photo. Sapiens2 supplies the dense supervision that makes a specific face possible: 308 keypoints across body, face, ears and hands (we use 272; the iris rows are excluded), a 29-class segmentation, surface normals, a metric point map and matting. In one local pass we get a skeleton, a mask, skin normals and depth — the measurements our solver is graded against.

Identity, pose and expression: SAM 3D Body on MHR
A SAM 3D Body forward pass gives a first estimate on the MHR base mesh. The released head, however, zeroes the 72 expression coefficients and jaw rotation — by design, that head is an identity/pose solver. 


Impact
Character creation stops being a modelling job. An indie studio can take one consented photo and get a riggable character with 72 face controls; a dubbing team can drive a performance without a scan; an accessibility tool can give someone an avatar that looks like them and can speak. The frame is fictional characters, not inference about real people — no medical, diagnostic or demographic claims.

## More from Meta Global AI Developer Hackathon

- [Physical prompting](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5180)
- [Si se puede: Porque somos panas](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5181)
- [Indie](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5182)
- [Ludavela](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5184)
- [Centaur Video](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5185)
- [Honeymeta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5186)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5183. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt