Stability AI and Tripo AI release image-to-3D AI model TripoSR
Stable Diffusion creator Stability AI and AI 3D modeling start-up Tripo AI have released TripoSR, a new open-source AI model for generating 3D models from single source images.
It can produce a “high-quality 3D model … in under a second”, and is “designed to cater to the growing demands of entertainment, gaming, industrial design, and architecture professionals”.
What does TripoSR do?
TripoSR generates a textured 3D mesh from a single source image, automatically generating the geometry and textures not visible in the original view.
The model was trained on a subset of the publicly available Objaverse research data set, using 3D models available under a Creative Commons CC BY license.
How does TripoSR compare to other image-to-3D reconstruction AI models?
In the technical report accompanying the release, Stability AI and Tripo AI compare TripoSR to other open-source 3D reconstruction models like OpenLRM and One-2-3-45.
It outperformed the other models tested for the accuracy of the models generated, and does pretty well for processing time.
So what does that mean in practice for 3D artists?
The results look pretty good in Tripo AI’s demo (at the top of this story), although the simpler models in the video from Stability AI’s blog post (above), may be more typical of the output.
Early user tests, such as this video from GamesFromScratch, suggest that the results for hard surface objects could be usable as background models for games, or for AR applications, although the results for organic characters are… interesting.
License and system requirements
The source code for TripoSR is available on GitHub under a MIT license, along with a list of dependencies. The model weights are available on Hugging Face.
At default settings, the model requires about 6GB of VRAM for a single image input, but according to Stability AI, it can be run on machines without GPUs.
Read an overview of open-source 3D generation model TripoSR in Stability AI’s blog post