# Escape This Photo

- **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:** [Guillermo del Rio](https://cerebralvalley.ai/u/brisaworks)
- **GitHub:** https://github.com/guillermo-delrio/escape-this-photo
- **Demo video:** https://youtu.be/KVO9DgZF2nQ
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3118

Escape This Photo turns a photo of a room into a point-and-click escape room you play on your phone, and checks every
room in code before anyone plays it, so every room has a way out.

Play it: https://escape-this-photo-74y6x74gra-uc.a.run.app
Start with Van Gogh's The Bedroom (written by hand, with outlines traced by SAM 3.1), then the gallery of rooms the
pipeline built from openly licensed photos. Each room has a "how it was made" page that shows every step below.

To build a room from your own photo, open https://escape-this-photo-74y6x74gra-uc.a.run.app/rooms?code=d6f38db5fded0011500ae685
(judges' access code: 40 rooms, until 10 Nov). A room takes about five minutes to build, six to seven with painted walls.

How a room is made:
1. Muse Spark checks the photo (no people, nothing unsafe) and lists up to 12 things in it a player could tap.
2. SAM 3.1 outlines each of them. The outlines are what you tap, and SAM's counts and sizes are the facts the puzzles
   must rest on: when a puzzle says three pictures, SAM counted three.
3. Muse Spark writes the puzzles: count real objects, read real text, find a small thing hidden in a bigger one, turn
   a fact into a lock code.
4. Code checks the model. Every number must match SAM's counts, every code must come from a fact the player can find,
   nothing needed may be covered by another object's outline, and a solver tries every move in every order to prove
   there is a way out and no way to get stuck. A room that fails goes back to Spark with the problems listed, up to
   twice; otherwise it is refused.
5. Muse Image paints what the photo can't show: the walls behind you, the inside of a drawer you open, the note you
   find. SAM measures the painted walls the same way it measures the photo.

Every gallery room was also played blind by an AI player built on Claude, from screenshots only, before it was shown.

Results on 15 openly licensed room photos (28 Oct): 12 became rooms. Two photos were refused for showing people, and
one failed its checks twice. Every build is in the repo's BENCHMARK.md. The count checks added after that run make each
room slower and dearer: in four builds on 4 Nov, one photo took about five minutes and $0.35, painted walls six to
seven minutes and $0.60, and three of the four became rooms.

A spend ledger books each model call's worst case before making it, with a daily cap and a fixed total, so the demo
stays up through judging without running past its credits.

## More from Meta Global AI Developer Hackathon

- [AegisQuorum](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3115)
- [ArkAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3116)
- [Persona B2B](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3117)
- [Vikzar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3119)
- [Ehtesham Warsi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3120)
- [Sawt](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3121)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3118. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt