TripoSR is an open-source model for fast, feedforward 3D reconstruction from a single image, developed collaboratively by Tripo AI and Stability AI. Where many 3D generation pipelines require minutes of per-object optimization, TripoSR produces a complete 3D model in under half a second on an NVIDIA A100 — a speed-quality combination that has made it a popular starting point for 3D generative AI, with more than 6,600 GitHub stars.

Single Image In, 3D Mesh Out

The workflow is deliberately simple: provide one image of an object, and TripoSR returns a 3D mesh. There is no need for multiple views, camera poses, or per-scene optimization. A single run.py command reconstructs the model and saves it to disk, and the project also ships a local Gradio app for an interactive, browser-based experience. For a single image, inference fits in about 6GB of VRAM, putting it within reach of consumer GPUs rather than only datacenter hardware.

Built on the Large Reconstruction Model

TripoSR is based on the principles of the Large Reconstruction Model (LRM), a transformer-based approach that maps image features directly to a 3D representation in a single forward pass. By avoiding iterative optimization, the model achieves its sub-second reconstruction times while, according to the authors' technical report, outperforming other open-source alternatives across multiple public datasets in both qualitative and quantitative evaluations. The architecture, training process, and comparisons are documented in an accompanying arXiv paper.

Textures and Output Options

Beyond raw geometry, TripoSR can produce usable assets. By default it outputs vertex colors, but the --bake-texture option generates a proper texture map, with --texture-resolution controlling the output resolution in pixels. Multiple images can be processed in one invocation, and the command-line flags expose the key knobs for batch and quality control. The result is a model that fits naturally into content-creation pipelines where a textured mesh, not just a point cloud, is what downstream tools expect.

Open and Permissive

One of TripoSR's biggest strengths is its licensing. The model — source code, pretrained weights, and an interactive online demo on HuggingFace Spaces — is released under the permissive MIT license. That makes it unusually friendly for both research and commercial experimentation, and it is one reason TripoSR is frequently used as a building block and baseline in newer image-to-3D projects. The stated goal of the collaboration is to empower researchers, developers, and creatives to push the boundaries of 3D generative AI and 3D content creation.

Considerations

TripoSR is a single-image reconstruction model, and that scope sets its limits. Quality depends heavily on the input: clean, well-lit images of a single foreground object work best, while cluttered scenes, occlusions, or unusual viewpoints can degrade results. As a fast feedforward model it trades some fidelity for speed, so the finest geometric detail and the back side of objects — which must be inferred from a single view — may not match slower multi-view or diffusion-based pipelines. Installation also requires matching local CUDA and PyTorch versions, which can trip up first-time users. For rapid prototyping, dataset generation, and any workflow that needs a quick 3D asset from a photo, however, TripoSR remains one of the most accessible and well-regarded open tools available.