Generating a 3D model from a 2D image used to be a slow, unreliable process. Most tools estimate depth in a rough "canonical space" and then guess at the rest, which is why results often look nothing like the source image. Pixal3D takes a different approach: it explicitly projects pixel features from the input image into a 3D volume, so the output stays faithful to what you actually gave it.
How it works
Built by researchers at Tsinghua University and Tencent ARC, and accepted to SIGGRAPH 2026, Pixal3D removes the spatial guesswork of canonical-space generators. Because every pixel feature is placed in 3D space with pixel-level alignment, the generated geometry and materials reach near-reconstruction-level fidelity to the input view. You get clean, detailed results instead of generic blobs.
What you can do with it
- Single or multi-view input: feed it one photo for fast generation, or several views for even better accuracy.
- PBR materials: output includes base color, normal, and roughness maps for realistic lighting in any renderer.
- GLB export: drop the result directly into game engines, Blender, or web viewers.
- Fully open source: MIT-licensed code you can self-host, extend, or study.
Why it matters
For game developers, e-commerce teams, and 3D hobbyists, the ability to convert existing 2D assets into production-ready 3D in seconds changes the workflow dramatically. Instead of hand-modeling every prop from scratch, you capture or find an image and let Pixal3D handle the rest. It is a practical, high-fidelity tool that makes image-to-3D generation feel less like a research demo and more like a daily utility.
Try it
The project is free and open source. You can run the model locally or check the live demo at pixal-3d.com — upload any image and watch it become a textured 3D mesh in seconds.
Sign in to leave a comment.