Introducing TripoSR: Fast 3D Object Generation from Single Images
Key Takeaways
Today we are releasing TripoSR in partnership with Tripo AI, generating high-quality 3D models from a single image in under a second.
TripoSR runs under low inference budgets (even without a GPU), which is both accessible and practical for a wide range of users and applications.
The model weights and source code are available for download here under the MIT license, allowing commercialized, personal, and research use.
We have partnered with Tripo AI to develop TripoSR, a fast 3D object reconstruction model inspired by the recent work of LRM: Large Reconstruction Model For Single Image to 3D. This new image-to-3D model is designed to cater to the growing demands of entertainment, gaming, industrial design, and architecture professionals, with responsive outputs for visualizing detailed 3D objects.
Sample input images and the corresponding 3D models were reconstructed using the TripoSR model.
Performance
TripoSR can create detailed 3D models in a fraction of the time of other models. When tested on an Nvidia A100, it generates draft-quality 3D outputs (textured meshes) in around 0.5 seconds, outperforming other open image-to-3D models such as OpenLRM. In addition to speed, our model is fully accessible to users with or without GPUs.
The plot shows the 3D performance with F-Score (higher the better) vs. inference time (lower the better).
Comparing TripoSR 3D reconstructions with those from OpenLRM.
Technical Details
Our training data preparation incorporates diverse data rendering techniques that more closely replicate the distribution of images found in the real world, significantly improving the model's ability to generalize. We carefully curated a CC-BY, a higher-quality subset of the Objaverse dataset, for the training data. On the model side, we also introduced several technical improvements over the base LRM model, including channel number optimization, mask supervision, and a more efficient crop rendering strategy. You can read the technical report for more details.
We invite developers, designers, and creators to explore its capabilities, contribute to its evolution, and discover its potential to transform their work and industries.
The code for the TripoSR model is now available on Tripo AI’s GitHub, and the model weights are available on Hugging Face. Please refer to our technical report for more details on the TripoSR model.
To stay updated on our progress, follow us on Twitter, Instagram, LinkedIn, and join our Discord Community.
