Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.livepeer.org/llms.txt

Use this file to discover all available pages before exploring further.

The Livepeer ComfyUI-Stream-Pack (livepeer/ComfyUI-Stream-Pack) is a suite of custom ComfyUI nodes for building real-time video and audio processing workflows with ComfyStream. Standard ComfyUI workflows operate on static images or pre-rendered video files. AI Stream Pack introduces node categories that accept live streams as first-class inputs, enabling the same node graph authoring experience for real-time pipelines.

​
Node Categories

The node pack uses three categories based on the role each node plays in a workflow:

​
Control Nodes

Control Nodes are the input/output layer of a streaming workflow. The current pack ships two Control Nodes:
  • LoadTensor: loads tensor data from a live stream for real-time processing. This is the entry point for video frames in a streaming workflow.
  • LoadAudioTensor: processes real-time audio input, delivering audio data into the workflow for audio-aware pipelines.
Both nodes replace the static Load Image pattern used in batch ComfyUI workflows. In a streaming context, they receive frames from the ComfyStream server on each forward pass instead of reading from disk.
Foundation Nodes and Light Nodes are defined in the node pack architecture but have no shipped implementations in the current release. The Control Nodes (LoadTensor, LoadAudioTensor) are the active entry points for streaming workflows.

​
Installation

The AI Stream Pack is installed as part of the ComfyStream setup. When ComfyStream is installed and its auxiliary custom nodes are copied into a ComfyUI workspace, the AI Stream Pack nodes become available in the ComfyUI node graph. To install standalone: 
cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/livepeer/ComfyUI-Stream-Pack.git
After restarting ComfyUI, the LoadTensor and LoadAudioTensor nodes appear in the node menu under the Livepeer category.

​
Using Stream Pack Nodes in a Workflow

A streaming workflow replaces the static Load Image node with LoadTensor. The rest of the workflow (samplers, ControlNet, VAE decode, etc.) is unchanged:
  1. Add a LoadTensor node as the input source
  2. Connect it to your img2img sampler or latent processing chain
  3. Export the workflow in API format
  4. Point ComfyStream at the exported JSON
ComfyStream’s server intercepts the LoadTensor node at runtime and injects decoded video frames. The downstream nodes process each frame as they would a static image input, but the forward pass is called once per incoming frame, not once per workflow execution. See for the full workflow setup and streaming server configuration. Stream Pack nodes work inside ComfyStream workflows. See the ComfyStream overview for the full real-time pipeline architecture.

Workflow Authoring

How to build and connect a ComfyUI workflow to a live video source with ComfyStream.

Real-Time AI Overview

Cascade architecture and how ComfyStream fits into the Livepeer pipeline.

ComfyStream as BYOC

Registering a ComfyStream instance with a Livepeer orchestrator.

PyTrickle Overview

Python SDK alternative for custom real-time processing without ComfyUI.
Last modified on May 19, 2026