Skip to main content
Page is under construction.

Check the github issues for ways to contribute! Or provide your feedback in this quick form
Livepeer AI Pipelines let you run customizable, composable video inference jobs across distributed GPU infrastructure. Powered by the Livepeer network and supported by off-chain workers like ComfyStream, the system makes it easy to deploy video AI at scale.

ā€‹
In a nutshell

  • Pipelines are one or more inference tasks (e.g. Whisper, style transfer, detection) run in sequence on video frames.
  • Gateways route jobs to compatible Orchestrators and workers; the protocol handles payment and coordination.
  • BYOC (Bring Your Own Compute) and ComfyStream are two ways to run or extend pipelines with your own models and nodes.

ā€‹
Use cases

  • Speech-to-text (Whisper)
  • Style transfer or filters (Stable Diffusion)
  • Object tracking and detection (YOLO)
  • Video segmentation (segment-anything)
  • Face redaction or blurring
  • BYOC (Bring Your Own Compute)

ā€‹
What is a pipeline?

An AI pipeline consists of one or more tasks executed in sequence on live video frames. Each task may:
  • Modify the video (e.g. add overlays)
  • Generate metadata (e.g. transcript, bounding boxes)
  • Relay results to another node
Livepeer handles stream ingest, frame extraction, and job dispatching. Nodes run the actual inference.

ā€‹
Architecture

ā€‹
Gateway and workers

  • Orchestrators queue inference jobs and run (or delegate to) workers.
  • Workers subscribe to task types (e.g. whisper-transcribe) and execute them.
  • Gateways route jobs from clients to compatible nodes. This is off-chain; the protocol (Arbitrum) handles payments and rewards.

ā€‹
Worker types

TypeDescriptionExample models

ā€‹
Pipeline definition format

Jobs can be JSON-based task objects. Example: 
{
  "streamId": "abc123",
  "task": "custom-pipeline",
  "pipeline": [
    { "task": "whisper-transcribe", "lang": "en" },
    { "task": "segment-blur", "target": "faces" }
  ]
}
Workers can accept:
  • JSON-formatted tasks via the Gateway
  • Frame-by-frame gRPC (low latency)
  • Result upload via webhook

ā€‹
Bring your own compute (BYOC)

You can use your own GPU nodes to serve inference tasks:
  1. Clone ComfyStream or implement the processing API.
  2. Add plugins for Whisper, ControlNet, or other models.
  3. Register your node with the gateway (and optionally on-chain).
See BYOC for a full setup guide.

ā€‹
See also

ā€‹
Resources

Last modified on February 18, 2026