ComfyUI Adds Native 3D Gaussian Splat Generation with TripoSplat

ComfyUI now turns a single image into a 3D Gaussian asset natively, through day zero support for TripoSplat, an open source image to gaussian splat model from Tripo (VAST AI) released under the MIT license. Updating ComfyUI to v0.23.0 exposes a TripoSplat workflow template in the Template Library. The template downloads the model, takes one uploaded image, and outputs a viewable Gaussian splat that drops into a modern 3D pipeline. Cloud and Desktop builds follow the v0.23.0 release.

Rather than spending the same number of Gaussians everywhere, the model concentrates geometry where the input image carries detail and stays lighter across simple regions, and it lets the user pick the detail level outright. Such as fewer Gaussians for background props, more for a hero asset, or several exported versions at different budgets in the manner of game level of detail.

The release is fully open. Model weights are published at huggingface.co/VAST-AI/TripoSplat and the inference code at github.com/VAST-AI-Research/TripoSplat, so the model runs locally and can be wired into custom workflows rather than sitting behind a service. Most commonly, people can expect to create 3D previews of stylized characters, environments, and props, 3D to 2D guidance, and AR/VR placement from a single concept image.

ComfyUI’s path into Gaussian Splatting predates this release, its earlier 3D support arrived through Comfy3D and VAST AI’s Triplane Gaussian, but native day-zero support for TripoSplat moves single-image splat generation into the shipped template set, with the model’s weights and code both open for local use. The downloadable workflow can be found on GitHub.