Install ComfyUI for Stable Diffusion

ComfyUI is a node-based interface for Stable Diffusion that gives you fine-grained control over every part of the inference pipeline. It runs well on Linux, but the setup involves Python virtual environments, CUDA or ROCm drivers, multi-gigabyte model files, and optional custom node extensions. This guide walks through a clean, repeatable installation from source.

Prerequisites and Driver Check

Before touching Python, confirm your GPU drivers and compute stack are in place. ComfyUI works with NVIDIA (CUDA), AMD (ROCm), Intel Arc (IPEX), and CPU-only, but NVIDIA on CUDA gives the best out-of-the-box experience.

NVIDIA — verify CUDA

nvidia-smi

You need driver ≥ 525 for CUDA 12. If nvidia-smi is missing, install the proprietary driver first. On Fedora/RHEL that means the RPM Fusion repo; on Ubuntu, ubuntu-drivers autoinstall or a manual PPA.

AMD — verify ROCm

rocminfo | grep -i 'agent type'

ROCm 5.7+ is required for most current PyTorch ROCm builds. Installation is beyond this guide's scope — see AMD's official ROCm docs for your distro.

System Packages

You need Python 3.10 or 3.11 (3.12 has occasional compatibility gaps with some custom nodes as of mid-2025), git, and a few build tools.

Debian / Ubuntu

sudo apt update
sudo apt install -y git python3.11 python3.11-venv python3.11-dev \
  libgl1 libglib2.0-0 build-essential

Fedora / RHEL 9 / Rocky 9

sudo dnf install -y git python3.11 python3.11-devel gcc gcc-c++ \
  mesa-libGL glib2

Arch Linux

sudo pacman -S --needed git python python-pip base-devel

Arch ships a current Python; no version pinning needed in most cases.

Clone ComfyUI and Create a Virtual Environment

Always run ComfyUI inside a venv. This isolates its dependencies from the rest of your system and makes upgrades clean.

git clone https://github.com/comfyanonymous/ComfyUI.git ~/ComfyUI
cd ~/ComfyUI

python3.11 -m venv .venv
source .venv/bin/activate

Your prompt will gain a (.venv) prefix. Every subsequent Python and pip command in this guide assumes the venv is active.

pip install --upgrade pip wheel

Install PyTorch

PyTorch must match your compute backend. Pick exactly one block below.

NVIDIA CUDA 12.1 (recommended for most NVIDIA GPUs)

pip install torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/cu121

NVIDIA CUDA 11.8 (older cards, e.g. GTX 10-series)

pip install torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/cu118

AMD ROCm 5.7

pip install torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/rocm5.7

CPU only (slow but universal)

pip install torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/cpu

Check PyTorch's official Get Started page if your CUDA version differs — the wheel URLs follow a predictable pattern.

Install ComfyUI Requirements

pip install -r requirements.txt

This installs safetensors, transformers, diffusers, accelerate, and other dependencies declared by the project. It may take several minutes on a slow connection.

Model Placement

ComfyUI expects models in specific subdirectories under models/. The structure mirrors the model type:

Model type	Directory
Stable Diffusion checkpoints (.safetensors, .ckpt)	models/checkpoints/
VAE	models/vae/
LoRA / LyCORIS	models/loras/
ControlNet	models/controlnet/
CLIP / text encoders	models/clip/
Upscalers (ESRGAN etc.)	models/upscale_models/
FLUX / SDXL UNET only	models/unet/

If your models already live elsewhere on disk, create symlinks rather than copying multi-gigabyte files:

ln -s /mnt/data/sd-models/v1-5-pruned-emaonly.safetensors \
  ~/ComfyUI/models/checkpoints/

Alternatively, edit extra_model_paths.yaml (copy the provided example and point it at your existing Automatic1111 or AUTOMATIC1111 model directories):

cp extra_model_paths.yaml.example extra_model_paths.yaml
nano extra_model_paths.yaml

Set base_path to your existing model root and ComfyUI will scan it on startup without any symlinking.

GPU Memory Flags

Stable Diffusion models range from 2 GB (SD 1.5 fp16) to 24+ GB (FLUX dev BF16). Tune memory behaviour with command-line flags when launching ComfyUI:

• --lowvram — offloads most of the model to RAM; slower but works on 4–6 GB VRAM cards.
• --medvram — moderate offloading; good for 8 GB.
• --medvram-sdxl — like --medvram but specifically helps SDXL on 8 GB.
• --cpu — run entirely on CPU (very slow; useful for debugging).
• --fp8_e4m3fn / --fp8_e5m2 — quantise weights to 8-bit; dramatically reduces VRAM with minor quality loss on supported hardware.
• --force-fp16 — force fp16 precision; helps on older NVIDIA cards that struggle with bf16.
• --disable-xformers — disable xFormers if you hit compatibility errors; ComfyUI will fall back to PyTorch's built-in attention.

xFormers is optional but can improve throughput on NVIDIA. Install it with:

pip install xformers

First Launch

cd ~/ComfyUI
source .venv/bin/activate
python main.py

For a card with 8 GB VRAM running SDXL:

python main.py --medvram-sdxl

ComfyUI starts a local web server, typically on http://127.0.0.1:8188. Open that URL in your browser. The default workflow loads automatically. If you need remote access (e.g., a headless server), add --listen 0.0.0.0 — but do this only on a trusted network since there is no built-in authentication.

Installing Custom Nodes

Custom nodes extend ComfyUI with new node types — upscalers, face detailers, prompt helpers, video support, and more. The recommended approach is ComfyUI-Manager, which adds an in-browser package manager.

cd ~/ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Manager.git

Restart ComfyUI. A Manager button appears in the top-right of the UI. From there you can search, install, and update nodes without touching the terminal. Each node may pull additional pip dependencies; ComfyUI-Manager installs them into the active venv automatically.

To install a node manually without the Manager:

cd ~/ComfyUI/custom_nodes
git clone https://github.com/example/SomeCustomNode.git
cd SomeCustomNode
pip install -r requirements.txt

Create a systemd User Service (Optional)

If you want ComfyUI to start on login without opening a terminal, create a user-level systemd service:

mkdir -p ~/.config/systemd/user
nano ~/.config/systemd/user/comfyui.service

[Unit]
Description=ComfyUI Stable Diffusion
After=graphical-session.target

[Service]
Type=simple
WorkingDirectory=%h/ComfyUI
ExecStart=%h/ComfyUI/.venv/bin/python main.py --medvram
Restart=on-failure
Environment=PATH=%h/ComfyUI/.venv/bin:/usr/local/bin:/usr/bin:/bin

[Install]
WantedBy=default.target

systemctl --user daemon-reload
systemctl --user enable --now comfyui.service

Adjust the ExecStart flags to match your GPU memory situation. Logs are available via journalctl --user -u comfyui -f.

Verification

1. Open http://127.0.0.1:8188 in a browser.
2. Confirm a checkpoint appears in the Load Checkpoint node's dropdown.
3. Click Queue Prompt. Watch the terminal for CUDA/ROCm device output — it will print which device is being used.
4. A completed image should appear in the browser output node within seconds to minutes depending on hardware.

Troubleshooting

"No module named 'torch'" after activating venv

You're likely in the wrong directory or the venv was not activated. Run source ~/ComfyUI/.venv/bin/activate explicitly, then python -c "import torch; print(torch.__version__)" to confirm.

CUDA out of memory

Add --lowvram or --medvram to the launch command. If still failing, try --fp8_e4m3fn for supported checkpoints, or reduce the image resolution in the workflow's empty latent node.

Black or green output images

Almost always a VAE mismatch. Some SDXL checkpoints need an external SDXL VAE. Place it in models/vae/ and wire a Load VAE node into your workflow explicitly rather than using the one baked into the checkpoint.

Custom node fails to load

Read the terminal output — missing pip packages are the most common cause. Activate the venv and run pip install -r custom_nodes/NodeName/requirements.txt manually. Some nodes also require matching versions of transformers or diffusers; check the node's README.

ROCm: "HIP error" or no GPU detected

Confirm your user is in the render and video groups (sudo usermod -aG render,video $USER then re-login). Also verify HSA_OVERRIDE_GFX_VERSION is set if your GPU is unofficially supported (e.g., RX 6600: export HSA_OVERRIDE_GFX_VERSION=10.3.0).