Skip to main content
ComfyStream is a modular AI inference engine that integrates with Livepeer’s Gateway Protocol to execute video frame pipelines on GPU-powered worker nodes. It extends ComfyUI with Livepeer-compatible gateway binding, real-time stream I/O, dynamic node graphs and plugin chaining, and overlay rendering and metadata export. For a high-level overview and DeepWiki, see the ComfyStream full guide.

​
Architecture overview

​
Node types in ComfyStream

Node typeDescriptionExample models
These are exposed as modules in nodes/*.py and can be chained in graph format.

​
Example pipeline: caption overlay

{
  "pipeline": [
    { "task": "whisper-transcribe" },
    { "task": "caption-overlay", "font": "Roboto" }
  ]
}
ComfyStream converts this to an internal computation graph (e.g. WhisperNode → TextOverlayNode → OutputStreamNode).

​
Plugin support

You can build your own plugins:
  • Implement the NodeBase class from ComfyUI
  • Register metadata and parameters
  • Declare inputs and outputs for chaining
Example: 
class FaceBlurNode(NodeBase):
  def run(self, frame):
    result = blur_faces(frame)
    return result

​
Connecting to Livepeer Gateway

In config.yaml: 
gatewayURL: wss://gateway.livepeer.org
models:
  - whisper
  - sdxl
Start your node: 
python run.py --adapter grpc --model whisper --gpu
The ComfyStream worker will listen to task queues via pub/sub, execute pipelines frame-by-frame, and return inference results as overlays or JSON.

​
Debugging pipelines

ComfyStream logs heartbeats to the gateway, job payloads, graph errors, and output stream metrics. Enable verbose mode: 
python run.py --debug

​
See also

​
Resources

Last modified on February 24, 2026