isaac-mcfadyen / docker-comfyui

The ComfyUI Stable Diffusion client packaged in a Docker container for easy redistribution.

comfyui-docker

The ComfyUI Stable Diffusion client packaged in a Docker container for easy redistribution.

• Only NVIDIA GPUs supported for now.
• Make sure to install the NVIDIA Container Toolkit and restart your Docker daemon before first use.
• View the example compose.yaml for instructions on how to persist things across reboot. WARNING: not persisting the /comfyui folder will cause every dependency to be redownloaded when the container is destroyed and recreated.

Configuration

• Based on my (extremely limited) testing, xformers can be faster on some older GPUs than PyTorch 2's Cross Attention. If you have an older GPU (RTX 2xxx, maybe some RTX 3xxx GPUs), make sure to set USE_XFORMERS to true when running the container. Newer NVIDIA GPUs should keep it set to false (the default) for greater speedups.
• Set COMFYUI_COMMIT to a git commit hash to pull a specific commit of ComfyUI, or to master (the default) to pull the latest version of ComfyUI.

Why can't we embed ComfyUI in the image directly? Why is it installed at build-time?
Some ComfyUI custom nodes require other dependencies. We could simply embed ComfyUI in the Docker image directly, but even if we bind-mounted the custom_nodes folder into the container the custom nodes would still need to reinstall their dependencies every run, which makes for quite a slow startup.