TripoSR

Open-source image-to-3D reconstruction model you can run locally

AI 3D & GamingFree planOverseasβ˜…β˜…β˜…β˜…β˜† 4.0

What is TripoSR?

TripoSR is an open-source feed-forward model, released by Stability AI and Tripo AI, that reconstructs a 3D mesh from a single image in well under a second on a modern GPU. Unlike hosted services it runs locally, so you control the pipeline, batch your own data and avoid per-generation fees. The repository ships inference code, pretrained weights and a Gradio demo, and output is a base mesh with textures or vertex colours depending on settings. It is a research-grade building block rather than a finished product, which is exactly what makes it useful for developers embedding reconstruction into their own tools.

Last updated: 2026-09-20. This site only provides an index; for exact features, pricing, and licensing, see the official website.

Key features

  • Single-image to 3D mesh reconstruction in a single forward pass
  • Runs locally on a GPU with released pretrained weights
  • Command-line inference plus a bundled Gradio demo UI
  • Exports OBJ meshes with textures or vertex colour
  • Batch processing over folders of input images
  • Python API for embedding reconstruction in custom pipelines

Pros & cons

Strengths

  • No per-generation cost once you have a GPU
  • Very fast inference compared with optimisation-based methods
  • Open weights allow fine-tuning and full pipeline control

Watch out for

  • Requires Python skills and a working GPU setup
  • Geometry from one image is approximate and needs cleanup
  • No support, hosting or user interface beyond the demo

Best for & use cases

developers embedding 3d reconstruction, research experiments and pipeline automation

If you're comparing similar products, check the alternatives below, or browse all tools in the AI 3D & Gaming category.

FAQ

What hardware does it need?

A CUDA-capable GPU is the practical requirement for reasonable speed, and the repository includes guidance on memory. CPU-only runs are possible but slow enough to be impractical for volume work.

Is the output production ready?

Not directly. It produces a plausible mesh from one view, so hidden geometry is guessed. Expect to clean topology and retexture before the model is used in a shipped product.

How does it differ from the hosted Tripo service?

The hosted service adds refinement passes, rigging and an interface. TripoSR is the open research model: fewer features, no fees, and full control when you run it yourself.