Chanjoong (CJ) Kim

Seoul

All projects

Realroom. Real furniture in a photo of your room.

When
Oct 2026 to Nov 2026
Stack
Meta Model API, Python, FastAPI, React, Modal
Two demo rooms, before and after, each beside its brief: a living room whose dark sofa became a beige one, with a rug and a floor lamp added, and an empty room furnished with a sofa, a rug, a coffee table, a floor lamp and a planter.
Two of the three demo rooms, both generated with Muse Image 1.0.

Why

If you are shopping for a sofa, the picture that helps is your own room with a real sofa in it, one that fits the space, and everything else left as it was. I wanted an agent that makes that picture.

How

Muse Spark 1.3 reads the room and plans what to add or replace, and SAM 3.1 outlines each object. For every spot the agent keeps only products of the right kind whose listed dimensions fit, out of a catalog of 4,375. It ranks them by how well their descriptions and photos match the plan, with Meta's Perception Encoder reading the photos. Muse Image 1.0 draws Spark's picks into the photo.

Each product is then checked against its catalog photo in three separate requests to Spark, and counts as matched only when at least two agree; one that came out wrong is redrawn. In the finished photo only the new products, their shadows and the spots that were cleared are drawn. Every other pixel is the original's.

The living-room run replayed in the Realroom app: the room's objects outlined, boxes where the sofa, the rug and the floor lamp go, three candidates drawn, each new product scored against its catalog photo, then a slider between the photo before and after.
The living-⁠room run, replayed in the app: the scan, the placement guide and the verify step, then a slider between before and after. Product photos: Amazon Berkeley Objects by Amazon.com, CC BY 4.0.

Outcome

I built Realroom alone for the Meta Global AI Developer Hackathon (the Indie Developer track, under the problem statement "Build for Your World") and submitted it with its code public on GitHub under the MIT license. On the three demo rooms every product was matched (3 of 3, 5 of 5 and 3 of 3), and the finished photos kept 68.7%, 63.6% and 67.5% of their pixels untouched. The runs took 433, 384 and 477 seconds and cost an estimated $0.39, $0.58 and $0.40 at Meta's published prices.