ComfyUI

AI Images & Video

Node-based image and video generation that runs on your own GPU.

ComfyUI is a node-based interface for running image and video generation models on your own GPU. You build a workflow by connecting nodes for loading a model, writing prompts, sampling and saving, so every step is visible and reusable. It supports Stable Diffusion and SDXL, newer image models, video models such as Wan, upscalers, ControlNet and a large ecosystem of custom nodes with a built-in manager. Workflows are saved as JSON and everything runs locally, so prompts and images stay on your network. It is a self-hosted alternative to paid image-generation services.

License

GPL-3.0

GitHub stars

141,955

Official website

https://www.comfy.org

Source code

https://github.com/Comfy-Org/ComfyUI

Compatibility

Architectures

amd64, arm64

Database

Stateless

Resource footprint

8192 MB RAM · 4 CPU

Measured on 2026-10-22 on ComfyUI 0.38.0 in a Proxmox LXC (4 vCPU, 14 GB RAM) with an RTX 3090 passed through. The service idles at about 1.7 GB of RAM. A 1024x1024 text-to-image run with RealVisXL V5 (SDXL, 25 steps) used about 7 GB of VRAM and about 8 GB of container RAM, including file cache. The first image took about 29 seconds because the model had to load from disk; the following ones took about 8 to 9 seconds. Video models and upscalers need much more: we cap the service at 13 GB of RAM after an uncapped heavy test froze the host. Plan for 100 GB of disk or more, because checkpoints are 6 GB and up.

Proof of Work Gallery

Workflow
Result
Resource usage

Deployment

# ComfyUI has no official Docker image. We run it natively in an LXC as a
# systemd service from a Python venv in /opt/ComfyUI (this is our real unit).
# /etc/systemd/system/comfyui.service
[Unit]
Description=ComfyUI
After=network.target

[Service]
User=root
WorkingDirectory=/opt/ComfyUI
ExecStart=/opt/ComfyUI/venv/bin/python /opt/ComfyUI/main.py --listen --port 8188 --disable-pinned-memory
Restart=on-failure
# Keep a runaway job from taking the whole host down.
MemoryMax=13G
MemorySwapMax=0

[Install]
WantedBy=multi-user.target

Alternative to

Midjourney, DALL-E, Runway