robertvoy/ComfyUI-Distributed

What it solves

It enables distributed and parallel processing for ComfyUI, allowing users to leverage multiple GPUs across a single machine, a local network, or cloud services to increase image and video generation throughput and accelerate upscaling workflows.

How it works

The extension introduces a master-worker architecture where a master ComfyUI instance coordinates tasks. It supports three types of workers: local (additional GPUs on the same machine), remote (GPUs on other computers in a network), and cloud (GPUs via services like Runpod). Users can convert workflows to distributed versions by adding specific nodes like Distributed Seed and Distributed Collector to handle task dispatching and result aggregation.

Who it’s for

Digital artists and AI researchers using ComfyUI who possess multiple GPUs or access to cloud GPU clusters and want to reduce the time spent on batch generation or high-resolution upscaling.

Highlights

• Parallel Workflow Processing: Run the same workflow with different seeds across multiple GPUs simultaneously to scale output.
• Distributed Upscaling: Accelerates the Ultimate SD Upscale by distributing image and video tiles across available workers.
• Flexible Worker Setup: Supports local, remote, and cloud-based GPU workers with auto-setup for local GPUs.
• Orchestrator-Only Mode: Allows the master machine to coordinate workers without performing any rendering itself.
• Developer API: Provides a POST endpoint to control the distributed cluster programmatically.