SwarmUI Documentation

Swarm (the name implies its ability to act as a 'swarm' of resources) allows you to connect multiple GPUs, multiple machines, or even remote machines over a network.

• Automatic Queuing: Generations are automatically queued across connected backends.
• Workflow Splitting: You can split a single custom ComfyUI workflow across multiple backend GPUs (subject to certain limitations).

For specific details on how to implement this, refer to the Using More GPUs documentation.

SwarmUI provides a variety of advanced features for image generation. Detailed documentation for specific features is organized into individual files within the /docs/Features/ directory. Key feature areas include:

• Presets: Configuration sets for generation.
• Prompt Syntax: Advanced text input features.
• Autocompletions: Prompt engine assistance.
• Image Prompting: Using IP-Adapter, ReVision, etc.
• ControlNet: Controlling image structure.
• Custom Comfy Workflows: Integrating custom ComfyUI workflows.
• Webhooks: Triggering external actions via SwarmUI.
• UISounds: Audio feedback for UI events.
• AutoScalingBackend: Advanced usage for Slurm/Kubernetes.
• Upscaling: Increasing resolution and quality for images and videos.

SwarmUI provides a full-capability network API for external programs to use Swarm features (like generating images) and manage the Swarm instance (like modifying backends).

Key Concepts:

• Request Format: Most API routes use POST requests sent to (your server)/API/(route) containing JSON inputs and receiving JSON outputs.
• Websockets: Routes with a WS suffix (e.g., GenerateText2ImageWS) use WebSocket connections. These typically take one initial input and provide multiple outputs over time, such as progress updates or preview images.
• Discovery: You can observe API calls by opening the SwarmUI web interface and using your browser's developer tools to inspect network traffic.

If an 'Other Machine' is intended to act as a backend for a Main Machine, it must be configured to accept LAN connections.

During Installation: Select the Just Yourself, with LAN access option in the installer UI.

Post-Installation Fix: If the server reports it is only accessible from the local computer, navigate to: Server Configuration -> Network -> Host and set the value to 0.0.0.0.

Click Save and restart SwarmUI.

Zeta Chroma is a pixel-space variant of Z-Image. Note that it is under active development and results may be poor; Z-Image LORAs are not compatible.

Setup:

• Download model files from lodestones/Zeta-Chroma.
• Save files in diffusion_models.
• This model does not use a VAE.

Recommended Parameters:

• CFG Scale: Normal ranges (e.g., 4 or 7).
• Steps: Normal counts (e.g., 20), but higher values may help compensate for the early pretrain.
• Sampler/Scheduler: Defaults are fine.
• Resolution: Standard is 1024. Supports a range from roughly 512 to 2048.

Use the <clear:texthere> tag to automatically remove parts of an image to make them transparent. This requires a PNG file format.

Example: <clear:background> will attempt to clear the background area.

The Self-Start method allows SwarmUI to configure, launch, and manage the ComfyUI backend automatically. This is the highly recommended installation method.

1. Ensure you have a valid ComfyUI installation. If you used the SwarmUI installer, it is located at dlbackend/comfy/ComfyUI/main.py.
2. In SwarmUI, navigate to Server -> Backends.
3. Click ComfyUI Self-Starting.
4. Set the StartScript path to your ComfyUI entry point (e.g., dlbackend/comfy/ComfyUI/main.py).
5. Save the backend configuration.

For webhooks that support image metadata (such as the Every Gen webhook), you can inject dynamic values into your JSON body using %tag% syntax. These values are JSON-escaped to ensure valid string formatting.

Common tags include:

• %prompt%: The prompt used for the generation.
• %image%: The URL to the generated image (formatted via the External URL setting in Server Configuration).

Warning: If using %image% with DoNotSave enabled, the URL may be a very large Base64 blob. Additionally, if your server is not externally accessible, external services like Discord will not be able to resolve the image URL.

{
  "username": "SwarmUI",
  "content": "Generated your image! Prompt was `%prompt%`, link is [here!](%image%)"
}

SwarmUI distributes tasks to backends based on their order in the backend list. The system follows a sequential queueing pattern:

• The first backend in the list receives the first generation in the queue.
• The second backend is only utilized if there are at least 2 generations queued simultaneously.
• The third backend is only used if there are at least 3 generations queued, and so on.

Best Practice: Arrange your backends in the list from fastest/best to slowest to ensure optimal performance.

To refine or upscale images:

1. Locate the Refine / Upscale parameter group (below Init Image) and toggle the checkbox to enable it.
2. For Refiners: Enable the Refiner Model parameter and select your model.
3. For Upscaling: Set the Refiner Upscale value (e.g., 1.5 or 2).
4. For DiT models (SD3/Flux): Enable Refiner Do Tiling.
5. Click Generate.