ComfyUI Extension: ComfyUI-Distributed
A custom node extension for ComfyUI that enables distributed image generation across multiple GPUs through a master-worker architecture.
Custom Nodes (0)
README
ComfyUI-Distributed
Supercharge your ComfyUI workflows with multi-GPU processing
A powerful extension for ComfyUI that enables parallel and distributed processing across multiple GPUs and machines. Speed up your image generation and upscaling workflows by leveraging all available GPU resources in your network.
Key Features
Parallel Workflow Processing
- Parallel Generation - Run the same workflow on multiple GPUs simultaneously with different seeds
- Automatic Load Balancing - Distribute workflow execution across available workers
- Batch Acceleration - Generate multiple variations faster by using all your GPUs
Distributed Upscaling
- True Distributed Processing - Split large upscaling tasks into tiles processed across multiple GPUs
- Tile-based Upscaling - Intelligent work distribution for Ultimate SD Upscale
Management & Monitoring
- Automatic Worker Management - Launch and monitor workers from the UI
- Network Support - Use GPUs across different machines on your network
- Real-time Monitoring - Track worker status and performance from the UI
- Easy Configuration - JSON-based configuration with UI controls
- Memory Management - Built-in VRAM clearing
Requirements
- ComfyUI installation
- Multiple GPUs
- That's it
Installation
-
Clone this repository into your ComfyUI custom nodes directory:
git clone https://github.com/robertvoy/ComfyUI-Distributed.git
-
Restart ComfyUI - If you'll be using remote workers, add
--enable-cors-header
to your launch arguments
Quick Start
Adding Local Workers
Local Workers: Additional ComfyUI instances running on the same computer (with multi-GPUs) as your main ComfyUI installation.
- Open the Distributed GPU panel.
- Click "Add Worker" in the UI.
- Configure your local worker:
- Name: A descriptive name for the worker (e.g., "My Gaming PC GPU 0")
- Port: A unique port number for this worker (e.g., 8189, 8190...).
- CUDA Device: The GPU index from
nvidia-smi
(e.g., 0, 1). - Extra Args: Optional ComfyUI arguments for this specific worker.
- Save and optionally launch the local worker.
Adding Remote Workers
Remote Workers: ComfyUI instances running on completely different computers on your network. These allow you to harness GPU power from other machines. Remote workers must be manually started on their respective computers and are connected via IP address.
- On the Remote Worker Machine:
- Launch ComfyUI with the
--listen --enable-cors-header
arguments. ⚠️ Required! - Add workers in the UI panel if the remote machine has more than one GPU.
- Make sure that they also have
--listen
set inExtra Args
. - Then launch them.
- Make sure that they also have
- Open the configured worker port(s) (e.g., 8189, 8190) in the remote worker's firewall.
- Launch ComfyUI with the
- On the Main Machine:
- Open the Distributed GPU panel (sidebar on the left).
- Click "Add Worker."
- Enable "Remote Worker" checkbox.
- Configure your remote worker:
- Name: A descriptive name for the worker (e.g., "Server Rack GPU 0")
- Host: The remote worker's IP address.
- Port: The port number used when launching ComfyUI on the remote worker (e.g., 8189).
- Save the remote worker configuration.
Services like Tailscale create secure VPN connections that make remote computers appear as if they're on your local network. This lets you use GPUs from:
- Other locations (work computer, friend's gaming PC, etc.)
- Cloud instances (AWS, Google Cloud, etc.)
- Any internet-connected machine with CUDA GPUs
Just install Tailscale on both machines and use the Tailscale IP address as the "Host" when configuring your remote worker.
</details>Configuration Tips
| Setting | Description | Example |
|---------|-------------|---------|
| CUDA Devices | Use nvidia-smi
to see GPU indices | 0, 1, 2... |
| Ports | Each worker needs a unique port | 8189, 8190... |
| Extra Args | Additional ComfyUI arguments | See below |
Common Extra Args:
--listen
- Required for remote workers--enable-cors-header
- Required if using remoter workers--lowvram
- For GPUs with less memory--highvram
- For high-end GPUs--reserve-vram 2
- Reserves 2GB of VRAM. Recommended for your primary/display GPU
Nodes
Distributed Collector
Collects and combines results from distributed processing
Usage: Place after any image generation node to enable distributed processing. Works automatically when workers are enabled.
Distributed Seed
Ensures unique seeds across distributed workers for varied generations
Usage: Connect to any seed input. Each worker automatically receives an offset seed to ensure randomisation. Alternatively, you can connect a seed node directly to this node's seed input, which will automatically handle seed offsetting across all workers.
Ultimate SD Upscale Distributed
Distributed version of Ultimate SD Upscale that processes tiles across multiple GPUs, making upscaling much faster
Usage:
- Upscale your image with a regular upscale model (ESRGAN, etc.)
- Feed the upscaled image into this node
- Configure tile settings
- Enable workers for fast processing
UI Features
Distributed GPU Panel
The control centre for your distributed setup:
| Feature | Description | |---------|-------------| | Worker Status | Real-time status indicators | | Launch/Stop | Control individual local workers | | Clear Memory | Free VRAM on all workers | | Interrupt | Stop current processing | | Worker Logs | View real-time logs |
Workflow Examples
Basic Distributed Generation
- Create your normal ComfyUI workflow
- Add Distributed Seed → connect to sampler's seed
- Add Distributed Collector → after VAE Decode
- Enable workers in the UI
- Run the workflow!
Distributed Upscaling
- Load your image
- Upscale with ESRGAN or similar
- Connect to Ultimate SD Upscale Distributed
- Configure tile settings
- Enable workers for faster processing
Troubleshooting
Common Issues
<details> <summary><b>Workers won't start</b></summary>- Check ports are not in use:
netstat -an | grep 8189
- Verify CUDA device exists:
nvidia-smi
- Check ComfyUI path in worker logs
- Worker was started outside the UI
- Stop the worker manually and use the UI to relaunch
- Ensure all remote workers have the same models available
- Check that custom nodes are installed on all remote workers
- Check firewall settings for required ports
- Verify master IP is accessible:
ping 192.168.1.100
- Ensure same ComfyUI version on all machines
- Ensure ComfyUI-Distributed is installed on remote workers
- Add
--enable-cors-header
to your launch argument, on both master and remote worker
- If you are running a worker in a Docker container (e.g., using
mmartial/comfyui-nvidia-docker
), you might encounter an error indicatingmain.py
cannot be found in/basedir
, preventing the worker from launching or finding custom nodes. - Solution: Add the following flags to the "Extra Args" field when configuring the worker in the ComfyUI-Distributed UI:
--base-dir /basedir --enable-cors-header --listen 0.0.0.0
This ensures that the base directory within the Docker container is correctly set, allowing ComfyUI to locate its necessary files and custom nodes.
UI Panel Settings
The Distributed GPU panel includes several configuration options in the Settings section:
Debug Mode
- Description: Enable detailed logging for troubleshooting and monitoring distributed operations
- Default: Disabled
- Usage: When enabled, detailed debug information is logged to the browser console, including worker status updates, job distribution details, and network communication logs
Auto-launch Workers on Startup
- Description: Automatically launch enabled local workers when ComfyUI starts
- Default: Disabled
- Usage: When enabled, any local workers that are marked as "enabled" will be automatically launched in the background when the ComfyUI server starts, eliminating the need to launch each worker manually
Stop Workers on Master Exit
- Description: Automatically stop all managed local workers when the master ComfyUI instance shuts down
- Default: Enabled
- Usage: When enabled, ensures clean shutdown by stopping all UI-managed worker processes when the main ComfyUI server exits, preventing orphaned background processes
Development
This project is under active development. Contributions are welcome!
Planned Features
- [x] Support for Cloud workers (release coming soon)
- [ ] Remote worker control via SSH
- [ ] View remote worker logs in UI
- [ ] Improve worker timeout logic
- [ ] ComfyUI Desktop app compatibility