vast-ai-research/triposplat

TripoSplat is a tool that takes a single photograph or image and converts it into a 3D representation. The 3D output uses a format called Gaussian Splatting, which represents a scene as a cloud of small blobs (called Gaussians) rather than as traditional polygons. This format is popular in real-time 3D viewing, game engines, and AR/VR because it can look very realistic while rendering quickly. The tool was created by TripoAI, a company focused on 3D AI research. You give it one image, and it produces a 3D file you can open in compatible viewers. The number of Gaussians in the output can be adjusted, up to about 262,000, letting you trade off between visual quality and how expensive the result is to display. The output files use the .ply or .splat formats, which are accepted by various 3D viewers. A notable aspect of the design is that the codebase is small, around 2,000 lines across two files, and avoids most heavy AI library dependencies. It does not require the large machine learning package collections that many similar tools rely on, which the authors describe as avoiding version-conflict problems. The model weights needed to run it are downloaded separately from HuggingFace or ModelScope. There are three ways to use it: a command-line inference script for basic use, a Gradio web interface you can run locally in your browser, and an integration with ComfyUI, a popular visual workflow tool for AI image and 3D work. The ComfyUI path requires only dropping in an official workflow template file. Setup involves installing Python, downloading the model weights (about a gigabyte or so), and installing a small set of standard packages. The code and model weights are released under the MIT license.