hoainho/img2threejs

img2threejs takes a single reference photo of an object and turns it into a working 3D model made entirely of code. Instead of extracting a mesh from the photo or downloading ready-made art, it writes a program that builds the shape from basic 3D building blocks, shaders, and generated geometry, so what you get is a live, animatable model rather than a static file. The tool is meant to be run inside an AI coding agent such as Claude Code, Codex, or OpenCode. Whichever agent is hosting it, img2threejs uses that agent's own way of looking at images, whether that is reading the picture directly, using a browser tool, or working from a screenshot someone provides. It handles both objects and characters differently. Plain objects, like a knife or a shotgun, go through a pipeline built for hard surfaces. Characters go through a separate track that pays attention to things like head proportions, facial features, and pose. Before it writes any code, the tool first lists out every small identifying detail in the photo, such as bevels, screws, engravings, or wear marks, and it will not move forward until that list is complete enough for the object's complexity. Building the model happens in a fixed sequence of passes, starting from a rough blockout and moving through structure, materials, surface detail, lighting, and final optimization. After each pass, the tool renders the model, takes a screenshot, and compares it against the reference image. It only continues to the next pass if that comparison scores well enough, and if a pass falls short, the tool automatically goes back and fixes either the plan or the code before trying again. The end result is a Three.js model with a real internal structure, meaning parts like pivots, sockets, and moving pieces are already set up, so the model is ready to animate rather than being an inert 3D shape.