ComfyUI's node spaghetti, served as an API

What it does

This is essentially a Docker wrapper that turns ComfyUI’s visual node editor into a remote JSON API. You export a workflow from the ComfyUI interface, POST it to a RunPod serverless endpoint, and receive generated images back as base64 strings or S3 URLs. Pre-built images bundle checkpoints for FLUX.1, SDXL, and Stable Diffusion 3, so you can skip the usual model-download chore.

The interesting bit

The clever part is the inversion: a tool famous for its browser-based node graph becomes a headless worker that only speaks JSON. It even accepts base64 input images and injects them into the workflow’s input directory before execution, letting you automate pipelines that were strictly point-and-click.

Key highlights

• Ships with pre-baked Docker images including FLUX.1 schnell/dev, SDXL, and SD3 checkpoints
• Returns results as base64 or uploads to S3 via environment variables
• Supports both synchronous (/runsync) and asynchronous (/run) endpoints
• Accepts base64-encoded input images by filename reference
• Version 5.0.0+ changed output to a structured list under output.images

Caveats

• RunPod imposes request size limits (10 MB for async, 20 MB for sync), so large base64 inputs can hit the wall
• Output format broke compatibility in v5.0.0; older integrations expecting output.message will need updating

Verdict

Good for teams that want to productionize ComfyUI pipelines without maintaining persistent GPU instances. Less useful if your workflow relies on heavy interactive tweaking, since you’ll still be exporting JSON from the GUI to test changes.