ComfyUI AMD Windows Setup

Install and run ComfyUI on AMD GPU Windows with DirectML — no CUDA needed.

ComfyUI is a powerful node-based Stable Diffusion workflow engine. This repository automates setup for AMD GPUs on Windows using DirectML backend, eliminating the need for NVIDIA-specific tooling or ROCm compilation.

Quick Start

git clone https://github.com/ChharithOeun/comfyui-amd-windows-setup.git
cd comfyui-amd-windows-setup

pip install -r requirements.txt

python scripts/install.py
python scripts/launch.py

Open your browser to http://localhost:8188 and start creating.

What is ComfyUI?

ComfyUI is a node-based interface for Stable Diffusion. Instead of typing prompts in a text box, you build workflows by connecting nodes:

• Model loaders → generators → samplers → output
• Easily branch pipelines for inpainting, upscaling, control networks
• Save workflows as JSON, run them repeatedly with tweaks
• Install custom nodes to extend functionality

Perfect for creative professionals, researchers, and automation workflows.

Features

✅ DirectML Backend — Full AMD GPU support on Windows
✅ Automatic Installation — One-command ComfyUI setup
✅ Custom Nodes — Easy install of popular node packs
✅ Model Management — Checkpoints, LoRAs, ControlNets in organized folders
✅ Cross-Platform Scripts — Works on Windows (batch + Python)
✅ AMD Optimized — --fp32 flag avoids black image issues
✅ Fallback to CPU — If GPU detection fails, runs on CPU

Usage

Launch ComfyUI

python scripts/launch.py

Optionally:

• --port 8080 — Run on custom port (default: 8188)
• --no-fp32 — Use fp16 (faster but may cause issues on some AMD GPUs)
• --cpu — Force CPU mode

Install Custom Nodes

python scripts/install_nodes.py --all

Or install a specific node:

python scripts/install_nodes.py --node ComfyUI-Manager

Verify GPU Setup

python scripts/verify_gpu.py

This checks DirectML support and diagnoses common issues.

Model Placement

Place models in these folders (created automatically):

• Checkpoints: ComfyUI/models/checkpoints/
• LoRAs: ComfyUI/models/loras/
• ControlNets: ComfyUI/models/controlnet/
• VAE: ComfyUI/models/vae/
• Inpainting: ComfyUI/models/inpaint/

Supported Models

• Stable Diffusion 1.5 — Original base model
• Stable Diffusion 2.x — Higher quality outputs
• SDXL — Latest, highest quality (requires more VRAM)
• ControlNet — Precise image guidance (pose, depth, canny edge)
• LoRA — Lightweight style/concept fine-tuning
• Inpainting Models — Mask-based image editing

Common Issues

No GPU Detected

Run python scripts/verify_gpu.py to diagnose:

• Are AMD drivers installed?
• Is torch-directml installed?
• Try --cpu mode as fallback.

Black/Pink Images (AMD fp16 Bug)

Solution: Use --fp32 flag (enabled by default). If explicitly disabled, add it back:

python scripts/launch.py  # includes --fp32 by default

Out of Memory

• Start with Stable Diffusion 1.5 instead of SDXL
• Reduce resolution: 512x512 instead of 768x768
• Use --fp16 if your GPU has sufficient VRAM and doesn't show artifacts
• Enable tiling or optimization nodes in ComfyUI UI

AMD-Specific Notes

Why --fp32?
Some AMD GPUs have issues with fp16 (half precision), resulting in black/pink outputs. --fp32 (full precision) is safer.

Disable xformers:
xformers is NVIDIA-optimized. ComfyUI auto-detects and skips it on DirectML.

Driver Compatibility:

• AMD Radeon drivers (20.50+) required
• Update via AMD Radeon Software or Windows Update

Part of AMD Windows AI Ecosystem

This repo is one of several tools for running modern AI on AMD Windows hardware:

• amd-windows-toolkit — Central hub for AMD GPU setup, LLM runners, environment validation
• comfyui-amd-windows-setup ← you are here
• Other models & runners coming soon

Check out the toolkit for unified AMD GPU discovery, PyTorch DirectML validation, and more.

Support

Found a bug or have a feature request? Open an Issue.

Want to contribute? See CONTRIBUTING.md.

License

MIT License — see LICENSE

---

If this project saved you hours of setup, consider buying me a coffee!