Skip to main content
Livepeer’s AI network routes inference jobs from your application to a distributed pool of GPU operators (orchestrators). You submit a job, a gateway routes it to a capable orchestrator, the orchestrator runs inference, and the result comes back. Three integration patterns are available depending on what you need:
  1. Standard API Pipelines - call a hosted endpoint, get a result. No infrastructure needed.
  2. ComfyStream - run ComfyUI-based workflows on live video frames in real time.
  3. BYOC (Bring Your Own Compute) - bring your own model container; Livepeer routes jobs to it.

​
Start here in 5 minutes

​
Choosing Your Integration Pattern

​
Standard API Pipelines

Standard pipelines are available via any Livepeer gateway that supports AI inference. Send a request with your model ID and parameters; get back a result.

​
Available Pipelines

​
Quick Example (text-to-image)

import { Livepeer } from "@livepeer/ai";

const livepeer = new Livepeer({
  httpBearer: process.env.LIVEPEER_GATEWAY_API_KEY,
});

const result = await livepeer.generate.textToImage({
  prompt: "A futuristic cityscape at night, neon lights, photorealistic",
  modelId: "SG161222/RealVisXL_V4.0_Lightning",  // fast, warm model
  width: 1024,
  height: 1024,
  numInferenceSteps: 6,   // Lightning model - keep low (4-8)
  guidanceScale: 1.5,      // Lightning model - keep 1.0-2.0
});

// result.imageResponse.images[0].url
Model selection matters. Lightning-suffix models (e.g. RealVisXL_V4.0_Lightning) are optimized for speed - use 4-8 inference steps and guidance scale 1.0-2.0. Standard SDXL models need 20-50 steps and guidance 7.0-9.0. Check available models and warm status before selecting.

​
Available Gateways for AI

GatewayEndpointAuthBest For
Livepeer Studiohttps://livepeer.studio/api/beta/generateAuthorization: Bearer <LIVEPEER_STUDIO_API_KEY>Production apps
Cloud SPEtools.livepeer.cloudProvider-definedDevelopment and experimentation
Self-hostedYour gateway URLAuthorization: Bearer <LIVEPEER_GATEWAY_API_KEY>Custom routing, private models
As of 02-March-2026, Studio AI uses https://livepeer.studio/api/beta/generate; for Cloud SPE-managed access, check tools.livepeer.cloud for current direct API endpoint and auth requirements.

​
ComfyStream

ComfyStream integrates ComfyUI with the Livepeer gateway protocol to run AI pipelines on live video frames in real time. It’s the foundation of real-time AI video products like Daydream. How it works:
  1. Video stream is ingested and split into frames
  2. Each frame is sent to a ComfyStream worker node
  3. The worker runs the ComfyUI workflow graph on the frame (style transfer, detection, etc.)
  4. The processed frame is returned and reassembled into an output stream
ComfyStream targets 15-30 FPS throughput with TensorRT-accelerated models achieving 10x+ performance over standard inference. Use ComfyStream for:
  • Real-time style transfer on live streams
  • Per-frame AI effects (depth estimation, face animation)
  • Interactive AI art with webcam input

ComfyStream Guide

Full ComfyStream architecture, node types, and integration guide.

​
BYOC (Bring Your Own Compute)

BYOC lets you bring a custom model container into the Livepeer AI network. Your container receives jobs routed by gateways, executes inference, and returns results - while Livepeer handles routing, payment, and coordination. BYOC is the right path when:
  • Your model is fine-tuned or proprietary (not available in the standard pipeline set)
  • You need a specific inference runtime (vLLM, TensorRT, custom Python)
  • You want Livepeer to provide the routing and payment layer for your compute
Container requirements:
  • Expose an HTTP endpoint implementing the Livepeer AI worker API
  • Accept job payloads matching the gateway’s protocol format
  • Return results in the expected schema

BYOC Setup Guide

How to build, register, and deploy a BYOC container on the Livepeer network.

​
How the Network Routes AI Jobs

The gateway selects the best orchestrator based on capability (does it have the requested model warm?), performance history, pricing, and latency. Warm models - models already loaded in GPU memory - return results significantly faster than cold models that need to load first.
Check tools.livepeer.cloud/ai/network-capabilities to see which models are currently warm on the network before choosing your model ID.

​
Next Steps

AI Inference Quickstart

Build your first AI integration in minutes.

API Reference

Full endpoint reference for all AI pipelines.

ComfyStream

Real-time per-frame AI video pipelines.

BYOC

Bring your own model container.
Last modified on March 9, 2026