Question 1

What is the TripoSR paper?

Accepted Answer

The TripoSR paper is a technical report titled "TripoSR: Fast 3D Object Reconstruction from a Single Image", published on arXiv as 2403.02151 on March 4, 2024. It introduces TripoSR, an open-source model that reconstructs a textured 3D mesh from a single input image in under 0.5 seconds using a feed-forward transformer architecture. The report documents the model design, the data processing and training improvements over prior work, and benchmark comparisons against other open-source single-image reconstruction methods. TripoSR was produced as a collaboration between Stability AI and Tripo AI and released under the MIT license.

Question 2

Who are the authors of TripoSR?

Accepted Answer

The TripoSR technical report lists ten authors: Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. They represent a collaboration between Stability AI and Tripo AI (the VAST AI Research team), pairing Stability AI's generative modeling work with Tripo AI's 3D reconstruction research. The full author list and their affiliations are available on the arXiv abstract page.

Question 3

When was the TripoSR paper published?

Accepted Answer

TripoSR was submitted to arXiv on March 4, 2024, under the identifier 2403.02151, in the Computer Vision and Pattern Recognition category (cs.CV). It was released as a technical report alongside the open-source code and pretrained weights, rather than through a traditional conference or journal, so the arXiv version is the canonical reference to cite.

Question 4

How does TripoSR work?

Accepted Answer

TripoSR uses a feed-forward transformer architecture built on the Large Reconstruction Model (LRM) design. Given a single RGB image, it predicts a 3D representation in one forward pass, with no per-object optimization and no additional views required, and outputs a textured 3D mesh in under 0.5 seconds on a GPU. The paper credits its quality gains to substantial improvements in data curation, model design, and training techniques compared with the original LRM approach.

Question 5

How fast is TripoSR, and what makes it fast?

Accepted Answer

TripoSR produces a 3D mesh from a single image in under 0.5 seconds on an NVIDIA GPU, according to the technical report. Because it is feed-forward, the runtime is roughly constant per image and does not grow with an iterative optimization loop. That is the key difference from slower, optimization-based 3D reconstruction methods, which can take minutes per object.

Question 6

Is the TripoSR paper and model open source?

Accepted Answer

Yes. TripoSR is released under the permissive MIT license, which allows both research and commercial use. The source code is published on GitHub at VAST-AI-Research/TripoSR and the pretrained weights on Hugging Face at stabilityai/TripoSR, so the results reported in the paper can be reproduced and the model can be run locally or integrated into other products.

Question 7

How does TripoSR compare to other 3D reconstruction models?

Accepted Answer

In the technical report, TripoSR is evaluated on public datasets and reported to outperform other open-source single-image 3D reconstruction alternatives, both quantitatively on standard reconstruction metrics and qualitatively. Its main practical advantage is combining that reconstruction quality with sub-second feed-forward inference and a permissive MIT license, which makes it straightforward to adopt in research and production.

Question 8

How do I cite TripoSR?

Accepted Answer

Cite the arXiv technical report, arXiv:2403.02151, using the BibTeX entry provided on this page. It lists all ten authors, the title "TripoSR: Fast 3D Object Reconstruction from a Single Image", and the year 2024. When you are referencing the model rather than the report, it also helps to link the official GitHub repository (VAST-AI-Research/TripoSR) and the Hugging Face model card (stabilityai/TripoSR).

TripoSR: Fast 3D Object Reconstruction from a Single Image

Abstract

Key contributions

Single image to 3D in under 0.5s

Transformer architecture built on LRM

Open source, MIT licensed

State of the art among open models

Authors

How to cite

Paper resources

arXiv abstract page

Source code on GitHub

Model on Hugging Face

TripoSR paper FAQ

See the paper in action