Juhana Kaarlehto has released ComfyUI-InfiniSplat 0.1.3, a custom node that runs the InfiniSplat single image gaussian splatting model inside ComfyUI.
ComfyUI's built-in route from one image to a splat has been TripoSplat, which arrived as a native workflow template in v0.23.0. The same release made splats a native data type, and InfiniSplat now feeds into the Preview Splat node that followed in v0.25.0. ComfyUI-InfiniSplat is a third party package, published to the Comfy Registry as infinisplat on October 9, and is not part of ComfyUI core.
InfiniSplat comes from the ZJU3DV group at Zhejiang University and was accepted to SIGGRAPH Asia 2026 (Journal Track), according to its arXiv listing. Most feed forward single image methods predict Gaussians from fixed pixel grid locations, which holds up near the input view but scatters primitives under large viewpoint shifts. InfiniSplat instead places sample points along the depth derived surface structure of the scene, then decodes each Gaussian from image features queried at those points. The authors call the result a surface aligned representation.
The node, InfiniSplat (Image to PLY), takes one RGB image or a batch and returns a model_3d output for the built-in Preview Splat node, with optional .ply saving. ComfyUI-InfiniSplat samples 1,500,000 Gaussians by default before a final floater filter, and assumes a 30 mm focal length (35 mm equivalent) unless GeoCalib estimation or intrinsics from a MoGe2 node replace it. A query_chunk_size default of 20,000, against upstream's 80,000, lowers peak decoder memory without reducing the sample count.
The node requires NVIDIA CUDA and ComfyUI 0.39.2 or newer. Its installer builds a separate Python 3.11, PyTorch 2.9 and CUDA 12.8 environment outside ComfyUI's own and pins a specific upstream commit. Testing covered Windows on an RTX 5090 Laptop GPU with 24 GB of VRAM, and Linux is untested. The depth guided InfiniSplat checkpoint is not exposed, because upstream's xFormers kernels fail on that GPU's SM120 architecture and the non-xFormers fallback fails as well.
The wrapper code is MIT licensed. InfiniSplat's code falls under the ZJU3DV Project Registration License, which is free for personal, research, educational and evaluation use and requires a free registration before any project use by a company, organization or other entity.
ComfyUI-InfiniSplat is available now on GitHub.
