ComfyUI Setup Guide - ajgillis04/GillisDockerDepot GitHub Wiki
ComfyUI generates and edits images on your own GPU. Here it's the image back end for OpenWebUI Setup Guide (WSL2 Docker), so you can ask for a picture in chat, or upload a photo and ask for changes.
Requires an NVIDIA GPU with the NVIDIA container toolkit (or Docker Desktop's GPU support on Windows/WSL 2).
compose/server2/comfyui.yaml:
services:
comfyui:
container_name: comfyui.${HOST_NAME}
image: yanwk/comfyui-boot:cu130-slim-v2
gpus: all
environment:
TZ: ${TZ}
CLI_ARGS: "--preview-method none"
OLLAMA_URL: http://ollama.${HOST_NAME}:11434
ports:
- "127.0.0.1:${COMFYUI_PORT}:8188" # no login, so this machine only
volumes:
- comfyui_models:/root/ComfyUI/models # Docker volume, fast disk
- ${DOCKERDIR}/comfyui/output:/root/ComfyUI/output
- ${DOCKERDIR}/comfyui/input:/root/ComfyUI/input
- ${DOCKERDIR}/comfyui/user:/root/ComfyUI/user
restart: always
volumes:
comfyui_models:
name: comfyui_models-
ComfyUI has no login, so its port is published on
127.0.0.1only. Open WebUI reaches it over the Docker network by name. -
Keep models in a Docker volume. On Windows/WSL 2, loading multi-gigabyte models from a Windows drive (
/mnt/d) takes minutes; a volume on the WSL disk is fast. - Pick an image that doesn't open public tunnels by default. Some all-in-one ComfyUI images start Cloudflare quick-tunnels unless you turn them off.
Download into the volume's checkpoints, diffusion_models, text_encoders and vae folders, matching what your workflow expects. Smaller distilled models generate in seconds on a 12 GB card. Editing models that take reference images (for example FLUX.2 Klein 4B) let chat edit uploaded photos; expect faces to come out close rather than exact, and quality to drop with repeated edits.
A 12 GB card can't hold a large language model and an image model at once. A small custom node can ask Ollama to unload its model before each image (Ollama's API: a request with "keep_alive": 0) and let it reload afterwards. Expect the first chat reply after an image to be a few seconds slower.
Admin Panel → Settings → Images: engine ComfyUI, URL http://comfyui.<host>:8188, then import your workflow JSON (saved from ComfyUI with Export (API)) and map its prompt, size and seed nodes.
-
CUDA out of memory: another app holds the GPU (Ollama, a Tdarr node transcoding). Unload or pause it. - Blank or black images: the workflow's model names don't match the files in the volume.
- Slow first image: the model is loading from disk; later images are faster.