ComfyUI, But You Don't Fight the CUDA Matrix

What it does

This repository builds and publishes Docker images for ComfyUI, the node-based Stable Diffusion GUI. It covers NVIDIA (CUDA 12.6 through 13.2), AMD (ROCm 6 and 7), Intel XPU, and even a CPU-only fallback. The images come in two flavors: slim (ComfyUI plus Manager and common dependencies) and MEGAPAK (an all-in-one bundle with development toolchains and dozens of pre-installed custom nodes).

The interesting bit

This is packaging work, plain and simple — and that is the entire selling point. ComfyUI’s real friction lives in dependency hell: matching PyTorch builds to GPU drivers, chasing custom node requirements, and rebuilding when the upstream toolchain shifts. The maintainer tracks that matrix explicitly, noting for example that ComfyUI’s performance library currently targets CUDA 13.0, which in turn drops support for Volta, Pascal, and Maxwell cards. You get a working environment without reading PyTorch build scripts.

Key highlights

• NVIDIA coverage spans CUDA 12.6, 13.0 (recommended), and 13.2, with a compatibility table showing which GPU architectures each supports.
• AMD ROCm images come in two variants: PyTorch’s faster builds versus AMD’s more comprehensive builds.
• slim images are beginner-friendly; MEGAPAK images bundle GCC, Python 3.13, and a pile of custom nodes for heavy users.
• Intel XPU and CPU-only tags exist for non-NVIDIA hardware.
• Licensed under the Mulan Public License v2, which is drafted in both Chinese and English.

Caveats

• CUDA 13.0 and 13.2 images do not support older NVIDIA architectures (Volta, Pascal, Maxwell); you need CUDA 12.6 for those.
• The README notes these limitations stem from PyTorch’s toolchain, not the NVIDIA CUDA Toolkit itself.

Verdict

Worth a look if you run ComfyUI anywhere other than your host OS, or if you have ever spent an afternoon reconciling torch versions with nvidia-smi. If you already have a working bare-metal install and no container preference, this adds little.