My 10-year-old son and I were building a Roblox game together. AI wrote the Lua code, generated concept art. But when we needed 3D models for the game, there was nothing usable.

AI can write code, generate images, compose music, make videos. But 3D? There's no "Midjourney for meshes" yet. So I built one.

The tech: TripoSR (open source by Stability AI + Tripo) deployed on a HuggingFace Space with a shared H200 GPU, wrapped in a Gradio web app. Upload any photo, it removes the background, runs the image through a vision transformer, predicts a 3D density grid, extracts the mesh via marching cubes, and gives you a downloadable .obj or .glb in under 5 seconds.

What works well: objects with clean silhouettes on contrasting backgrounds. A green Hot Wheels on a white surface came out great. A grey truck on grey carpet, not so much.

What it's NOT: this isn't replacing Blender or ZBrush. Vertex-colored meshes, no UV-mapped textures (that needs 16GB+ VRAM for baking). 256 marching cubes grid resolution. But for quick prototyping, reference meshes, or game jam assets, it gets the job done.

Stack: Python, Gradio, PyTorch, HuggingFace Spaces (ZeroGPU).

Live at https://meshcraft.xyz - beta access is $9.97/mo (unlimited generations). Happy to answer technical questions.

Solo dev, 20+ years experience. Idea to deployed product in about 48 hours.