Filip Ziemski

CodeMesh 3D turns a phone video into a clean, game-ready 3D model. Walk around any object with your phone, upload the video, and type what it is ("electric kettle"). CodeMesh 3D does the rest automatically: - SAM 3.1 finds and tracks the object through every frame from that one sentence. No manual masking. - MapAnything (Meta open weights) recovers every camera position and depth map in one pass, on a Modal GPU. - Muse Spark 1.3 is the modeling agent. It studies the photos, picks how to build the object (for example a lathe-turned body plus a swept handle for a kettle or mug), and builds a clean mesh with tools and code. Every change is checked against the real footage, and the best version is kept. - A quality gate checks that the model is one watertight piece, grades it, and gives recapture tips if needed. Out comes a GLB with UVs, a texture and a normal map, ready for Blender, Unity, Unreal or a video editor. Our kettle result matches the real silhouettes at about 89%, at roughly a cent of API cost per job. Live: https://fjziemski--codemesh-web.us-east.modal.run Code: https://github.com/PenguinX156/codemesh-3d