Stability AI Releases 3D Generation Model TripoSR Capable of Producing High-Quality 3D Models in Under 1 Second
-
Stability AI and Tripo AI jointly released a 3D generation model named TripoSR last night. This model can generate high-quality 3D models in less than 1 second, marking a revolutionary advancement in the field of 3D modeling.
TripoSR requires minimal computational power for inference and doesn't even need a GPU, significantly reducing production costs. Additionally, the model's weights are licensed for commercial use, which is great news for many businesses.
Address: https://stability.ai/news/triposr-3d-generation In terms of performance, TripoSR can create detailed 3D models in a fraction of the time required by other models. When tested on an Nvidia A100, it can generate preliminary quality 3D outputs (textured meshes) in approximately 0.5 seconds, outperforming other open image-to-3D models such as OpenLRM.
In terms of technical details, the training data preparation for TripoSR includes various data rendering techniques, which bring the model closer to the distribution of real-world images and significantly improve the model's generalization capabilities. Additionally, they carefully curated a high-quality subset of the CC-BY Objaverse dataset for training data. On the model side, several technical improvements were made to the base LRM model, including channel optimization, mask supervision, and more efficient cropping and rendering strategies.
Overall, the collaboration between Stability AI and Tripo AI has resulted in the TripoSR 3D generation model, which not only achieves technical breakthroughs but also brings new possibilities to the field of 3D modeling.