TripoSplat is an open-source feed-forward model from VAST-AI Research (the lab behind TripoSR and TripoSG, and the team operating Tripo3D) that converts a single 2D image into 3D Gaussians in one forward pass. There is no per-scene optimization loop: the image goes in, a set of Gaussians comes out, and the result exports as .ply or .splat for any standard 3D Gaussian viewer such as SparkJS or SuperSplat.

The distinguishing design choice is the variable Gaussian count. Most single-image-to-3D-Gaussian models emit a fixed budget of primitives, which forces the same rendering cost onto a simple prop and a detailed character alike. TripoSplat lets the caller request an arbitrary number up to 262,144, making the quality-versus-rendering-cost tradeoff an explicit runtime parameter rather than a property baked into the checkpoint. A game asset destined for a mobile target and a hero asset for a cinematic can come from the same model at different budgets.

The engineering posture is unusually restrained for a research release. The core is two files — triposplat.py and model.py — totaling roughly 2,000 lines, and the runtime dependencies are numpy, safetensors, pillow, and tqdm on top of PyTorch. There is no transformers and no diffusers in the path, which sidesteps the version-pinning conflicts that make many 3D generation repos difficult to install alongside anything else. That compactness is a deliberate integration argument: the code is small enough to read end to end and graft into an existing pipeline.

Distribution is broad for a project of this size. Weights are published on Hugging Face (VAST-AI/TripoSplat) and mirrored on ModelScope, there is a Gradio demo (run_gradio.py) and a hosted Hugging Face Space, and ComfyUI ships an official workflow template for image-to-Gaussian-splat, so ComfyUI users can run it without writing glue code. The accompanying paper is on arXiv, with a technical blog on the Tripo3D site.

Two caveats are worth stating. The repository has seen no commits since early June 2026, so this is a stable drop rather than an actively iterated project — fine for a self-contained inference model, less reassuring if issues accumulate. And the output is a Gaussian splat, not a mesh: splats render beautifully and suit AR/VR, previsualization, and simulation backdrops, but a conventional game or DCC pipeline that needs watertight topology, UVs, and PBR materials still requires a separate meshing step. Everything is MIT licensed.