ComfyUI
ComfyUI is Comfy Org's open source node graph engine for generative models, with native nodes to preview, render and export Gaussian splats.
Platform
Desktop app (Windows, macOS), portable Windows builds, manual install on Windows, Linux and macOS, and Comfy Cloud in the browser
GPU
No single floor published. ComfyUI runs on NVIDIA, AMD, Intel and Apple Silicon GPUs, Ascend NPUs or the CPU, and Comfy Org states the largest open models run in as little as 4 GB of VRAM with 8 GB of RAM. The standard Windows portable NVIDIA build supports 20 series and newer, with a separate CUDA 12.6 portable build for 10 series and older cards. Comfy Cloud needs no local GPU
Pricing
Free and open source to run locally. Comfy Cloud is a paid hosted version
License
GPL-3.0
Compare licence terms ↗
Best for
Node based generative workflows that create, render and convert Gaussian splats
Updated
July 2026
ComfyUI is an open source node graph application for running generative models locally, built by Comfy Org and licensed under GPL 3.0. The project began on GitHub in January 2023 as a graph and node interface for diffusion models and has grown into a general creation engine for images, video, audio and 3D, where every model, parameter and output is a node that can be wired into a larger workflow.
ComfyUI runs on Windows, macOS and Linux. Comfy Org ships a desktop application for Windows and macOS, portable Windows builds for NVIDIA, AMD and Intel GPUs, and a manual install that supports NVIDIA, AMD, Intel, Apple Silicon and Ascend hardware. Comfy Org states that its asynchronous weight streaming runs the largest open models on as little as 4 GB of VRAM and 8 GB of system RAM. Comfy Cloud is the paid hosted version for users without local hardware, and partner nodes reach closed source models through the same graph.
Gaussian splatting first reached ComfyUI through community node packs. Comfy3D by Mr. For Example arrived in February 2024 with VAST AI's Triplane Gaussian image to splat model and in graph previews of 3DGS scenes and meshes. ComfyUI v0.23.0 then brought day zero support for TripoSplat, an MIT licensed image to Gaussian splat model from Tripo, as a workflow template that downloads the model, takes one image and returns a viewable splat at a detail level the user chooses.
The same release made splats a native data type. ComfyUI's SPLAT type stores positions, scales, rotations, opacities and spherical harmonic coefficients and flows between nodes the way images, masks and latents do. RenderSplat rasterizes color, clay, depth and normal passes with turntable orbits, SplatToMesh extracts a vertex colored mesh, TransformSplat and MergeSplat handle transforms and concatenation, and GetSplatCount reports how many splats a batch holds. The Get Splat loader reads .ply, .splat, .ksplat and .spz and can detect the format from the file contents, and the exporter writes .ply with full spherical harmonics, or .ksplat and .spz with base color only. A dedicated Preview Splat node for interactive previews inside the editor followed in v0.25.0.
Because the type is shared, any core or custom node can produce or consume a splat. A scene reconstructed in another tool can be loaded as a .ply, repositioned, rendered to depth and normal maps that guide a diffusion pass, and exported back to .spz for delivery inside a single graph with no Python scripting.
ComfyUI is available on GitHub.
Frequently Asked Questions
Does ComfyUI support Gaussian splatting natively?
Yes. ComfyUI has supported Gaussian splats natively since v0.23.0, which added a SPLAT data type and first party nodes to render, mesh, transform, merge, import and export splats, reading .ply, .splat, .ksplat and .spz and writing .ply, .ksplat and .spz. A Preview Splat node for interactive previews followed in v0.25.0.
Can ComfyUI generate a Gaussian splat from a single image?
Yes. ComfyUI ships a TripoSplat workflow template from v0.23.0 that turns one image into a Gaussian splat, and community node packs add further image to splat models.
What hardware does ComfyUI need?
ComfyUI runs on NVIDIA, AMD, Intel and Apple Silicon GPUs, Ascend NPUs and the CPU, and Comfy Org states the largest open models run in as little as 4 GB of VRAM with 8 GB of RAM. Comfy Cloud is a paid hosted option that needs no local GPU.
Is ComfyUI free?
ComfyUI is free and open source under GPL 3.0 when run locally through the desktop app, a portable build or a manual install. Comfy Cloud, the hosted version, is paid.


