TripoSR

TripoSR is an open-source 3D reconstruction model developed through a collaboration between Stability AI and Tripo AI. It is designed to generate high-quality 3D meshes from a single RGB image with extreme speed, often producing results in under 0.5 seconds on a standard GPU. The model is released under the MIT license, making it accessible for researchers, developers, and artists working in gaming, industrial design, and specialized 3D content creation.

The model's architecture is built upon the Large Reconstruction Model (LRM) framework and utilizes a transformer-based feed-forward approach. It consists of three primary components: a pre-trained vision transformer (DINOv1) used as an image encoder, an image-to-triplane decoder that maps latent vectors onto a triplane-NeRF representation, and a neural radiance field (NeRF) model comprised of multilayer perceptrons (MLPs). A key technical advancement in TripoSR is its ability to "guess" camera parameters during inference, which significantly improves robustness when handling real-world images that lack precise camera metadata.

TripoSR was trained on a curated subset of the Objaverse dataset using optimized rendering techniques that emphasize foreground surface details. To achieve an optimal balance between quality and performance, the training process employs random patch supervision at high resolutions and an importance sampling strategy. For the best generation results, it is recommended to provide input images with 1024x1024 resolution, clear subject focus, and minimal background obstruction.