TripoSR

TripoSR is an open-source 3D reconstruction model built by Tripo AI and Stability AI — not Microsoft. It turns a single image into a 3D model in under half a second on an NVIDIA A100 GPU. The design follows Large Reconstruction Model principles with a transformer architecture. It targets researchers, developers, and creatives building image-to-3D pipelines. Code, weights, and a live demo are MIT-licensed. Default inference needs about 6 GB of VRAM. A Gradio app and command-line tool are both included. Check the GitHub repo for setup steps.

Key Features

1. Half-Second Reconstruction

Feedforward design rebuilds a 3D model from one image in under 0.5 seconds on an A100. No per-sample optimization runs.

2. LRM-Style Architecture

Built on Large Reconstruction Model principles with a transformer backbone. It beats older open alternatives on public benchmarks.

3. Modest VRAM Needs

Single-image inference uses about 6 GB of VRAM by default. A CPU fallback exists through torchmcubes.

4. Texture Baking Option

Default output uses vertex colors. The --bake-texture flag exports real texture maps with adjustable resolution.

5. CLI Batch Inference

Run one command with an input image and output folder. Multiple images process in a single batch run.

6. Gradio and Live Demo

A local Gradio app ships in the repo. A hosted Hugging Face demo lets you try it with no install.

Frequently Asked Questions (FAQ)

Did Microsoft make TripoSR?

No. The README states it was collaboratively developed by Tripo AI and Stability AI. Microsoft is not involved.

Textures or vertex colors?

Vertex colors by default. Add --bake-texture to export texture maps instead.

What GPU do I need?

About 6 GB of VRAM for default inference. The 0.5-second claim was measured on an NVIDIA A100.

Can I try it without installing?

Yes. The Hugging Face Space demo runs it in the browser.

What are common setup errors?

torchmcubes CUDA mismatches top the list. The README covers matching PyTorch and CUDA versions plus setuptools requirements.

Conclusion

TripoSR is a fast, MIT-licensed image-to-3D model from Tripo AI and Stability AI. Sub-second reconstruction and modest VRAM needs suit research and creative pipelines. It is not a Microsoft product despite the name confusion. Start with the hosted demo, then run locally for batch work. Check the GitHub repo for code, weights, and the live demo.