Skip to Main Content

worldoverlay

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

worldoverlay — Demo video

WORLD OVERLAY is an AI assembly coach for people who have never built anything. Point your phone at a pile of parts and it walks you through one step at a time: which part to pick up, where it goes, and what the result should look like. The problem: furniture and kit manuals are abstract line drawings. Beginners can't match the drawing to the real parts on their table, so they give up or assemble things wrong. How it works (all in the phone browser, no app install): 1. Record one short video sweeping across all the parts (and optionally snap the manual). 2. Muse Spark watches the video, finds each part and its clearest frame, and plans the assembly order, reading the manual when provided. 3. SAM 3.1 cuts out every part from several prompts (box, box+points, text), and Muse Spark acts as a judge to pick the clean cutout with no table or hands. 4. Muse Image generates a photo of each step's result, consistent from step to step, so users see exactly what "done" looks like. 5. An animated guide flies part A onto part B, then reveals the generated result. 6. Live identification: point the camera at any part and Meta's open-weights DINOv2 recognizes it locally (about 0.4 s on CPU, no API cost), telling you "this is X, it connects to Y in step k". Muse Spark breaks ties. 7. Optional 3D view of each part with AR. The first step plays about 45 seconds after upload, and the rest stream in while you build. The pipeline works on any physical assembly task: furniture, kits, electronics, repairs, and factory onboarding. Built with Meta Model API (Muse Spark, SAM 3.1, Muse Image) and DINOv2. Live demo: https://fast-bobby-fewer-isle.trycloudflare.com Judge passcode: 706c9008f7ce (10 analyses per day; the app asks before each run. Need more? Email lai2570@gmail.com)

Team