# Latent Raw

- **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:** [Caio Silva](https://cerebralvalley.ai/u/Kyoos), [Swarnabha Ghosh](https://cerebralvalley.ai/u/SamGomes7125)
- **GitHub:** https://github.com/Atikum-AI/latent-judge-hackathon
- **Demo video:** https://youtu.be/WNvkfJxQJj0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3302

Latent Judge is a quality gate for AI video: it catches a bad render at an early step, before you pay for the rest.

Generative video fails silently. A night shot comes out pure black, a sunny street clips to white, and you only find out after paying for every step. But a diffusion model builds the whole clip at every step, so a rough draft exists long before the end. Latent Judge reads that early draft, predicts what the finished clip will measure (crushed blacks, clipped highlights, exposure, saturation, colour balance) and decides: continue, stop, or steer with one correction. Every decision is logged with the numbers it was made on.

Meta models, each with a job:
- Muse Spark is the director: it reads each shot, chooses what it should be judged against, and when the judge stops a draft it picks one correction from a closed menu and must give a reason. Every call is logged with its response id.
- SAM 3.1 tracks each person, hand and object through a clip, so a shot is judged per object, not only per frame.
- Muse Spark Open reads the run history and writes where the model fails by kind of light, and what to train next, citing the evidence.

Results (held out, pre-registered: 30 prompts Spark wrote blind, pass marks frozen in git before any GPU time):
- 0 good drafts wrongly stopped
- 25 of 30 delivered and passing with the judge, 18 of 30 without
- failures caught by step 3: 3 of 12 (our 50% mark: not met); by step 5: 8 of 12
- night porch shot: 54% crushed blacks predicted at step 3 against a 20% limit; left alone it finished at 76.4%; one shadow-lift correction delivered it at 0.01%
- on LTX-2.5's 50-step base model (calibration, not yet blind): 15 of 23 night failures caught by step 12 with no false stops, so 38 of 50 steps never run

Known limit: today the director's thinking time costs more than the GPU the gate saves; we already cut its calls from 45 to 9.

Next: a Gate API in private beta for platforms that pay for every wasted render, then a node for creator tools. Free audit: send us 50 failed generations and we show which ones we would have stopped.

All footage is generated; detail added in dark areas is synthesized, never recovered.

## More from Meta Global AI Developer Hackathon

- [Reel Talk](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3299)
- [Nzangi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3300)
- [PeerCodeAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3301)
- [T-1777](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3303)
- [Omukuumi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3304)
- [Creative WebMatrix Solutions](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3305)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3302. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt