Stability AI has announced TripoSR, a collaboration with Tripo AI. This new technology promises to generate 3D models from single images, achieving results in less than a second. TripoSR is designed to operate efficiently under low inference budgets, making it accessible to a broad spectrum of users across various fields, including entertainment, gaming, industrial design, and architecture.
The technology behind TripoSR draws inspiration from the Large Reconstruction Model For Single Image to 3D (LRM) project, showcasing impressive speed and quality in creating 3D models. Unlike other models, TripoSR is capable of generating detailed, draft-quality textured meshes in approximately 0.5 seconds on an Nvidia A100. Importantly, the model is designed to be fully accessible, functioning with or without the support of GPU hardware.
Technical advancements in TripoSR include diverse data rendering techniques for training data preparation, optimizing the model’s ability to generalize to real-world images. The development team utilized a curated subset of the Objaverse dataset, alongside improvements such as channel number optimization and a more efficient crop rendering strategy. The availability of TripoSR’s model weights and source code under the MIT license on Tripo AI’s GitHub and Hugging Face platforms invites widespread use and contribution from the development community.
For more details on TripoSR and to access the model and its technical report, visit Stabiltity AI’s website.
