Pixal3D is an image-to-3D generation model, a neural network that takes a picture of an object and produces a 3D model of it. It is a research project by Tsinghua University, Tencent ARC Lab and Victoria University of Wellington, accepted to SIGGRAPH 2026, with code published on GitHub.
Earlier methods pass image features to the 3D generator loosely, through attention. Pixal3D instead lifts each pixel's features into 3D space by back-projection (tracing the pixel back along the camera ray), so every pixel has a direct match in 3D. The result has detailed geometry and PBR textures (physically based rendering materials, which describe colour, roughness and metal for realistic lighting). A Python script turns one image into a GLB mesh file; a low-VRAM mode lowers the GPU memory needed, and a multi-view mode uses several photos of the same object. The model works in three stages, sparse structure, shape and texture. The main branch builds on the TRELLIS.2 model; a paper branch reproduces the published results.
Pixal3D is free and open source under the MIT license, with a browser demo on Hugging Face.