AI Jobs Direct Quickstart

Make your first AI inference call against the Livepeer network in under ten minutes. Free community gateway, no API key required for experimentation.

By the end of this quickstart you’ll have run a text-to-image inference call against the Livepeer network, downloaded the generated image, and seen the same call work through both @livepeer/ai (TypeScript) and livepeer-ai-python. The path needs only a terminal or a runtime.

The path runs against dream-gateway.livepeer.cloud, the free Cloud SPE Community Gateway. It is off-chain, public, and shaped for experimentation. Production workloads should run against a paid gateway or a self-hosted one; see for the trade-offs.

Required Tools

Pick one of the three paths below. You need exactly one:

• curl (any modern shell)
• Node.js 20 or later, with npm or pnpm
• Python 3.11 or later, with pip

The network does the GPU work. Your machine only sends the request and receives a URL.

First Call

1. Send a POST request

   The community gateway accepts unauthenticated POSTs at /text-to-image. Pick your client below.

   curl

   curl -X POST https://dream-gateway.livepeer.cloud/text-to-image \
     -H "Content-Type: application/json" \
     -d '{
       "model_id": "ByteDance/SDXL-Lightning",
       "prompt": "A coffee shop at sunset, photorealistic",
       "width": 1024,
       "height": 1024
     }'
   

   Node (fetch)

   const res = await fetch('https://dream-gateway.livepeer.cloud/text-to-image', {
     method: 'POST',
     headers: { 'Content-Type': 'application/json' },
     body: JSON.stringify({
       model_id: 'ByteDance/SDXL-Lightning',
       prompt: 'A coffee shop at sunset, photorealistic',
       width: 1024,
       height: 1024,
     }),
   });
   const data = await res.json();
   console.log(data.images[0].url);
   

   Python (requests)

   import requests
   
   res = requests.post(
       "https://dream-gateway.livepeer.cloud/text-to-image",
       json={
           "model_id": "ByteDance/SDXL-Lightning",
           "prompt": "A coffee shop at sunset, photorealistic",
           "width": 1024,
           "height": 1024,
       },
   )
   print(res.json()["images"][0]["url"])
   
2. Inspect the response

   The gateway returns a JSON object with an images array. Each image has a hosted URL, a deterministic seed, and an NSFW classifier flag.

   {
     "images": [
       {
         "url": "https://dream-gateway.livepeer.cloud/stream/d0fc1fc6/8fdf5a94.png",
         "seed": 2562822894,
         "nsfw": false
       }
     ]
   }
   

   The url is publicly fetchable.

3. Download the image

   curl -O https://dream-gateway.livepeer.cloud/stream/d0fc1fc6/8fdf5a94.png
   

   Open the downloaded .png. That image was generated on a GPU run by an independent orchestrator on the Livepeer network.

SDK Path

The same call works through the official SDKs. Pick the language that matches your project.

@livepeer/ai (TypeScript)

Install the package:

npm install @livepeer/ai

Call the gateway:

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

const livepeer = new Livepeer({
  serverURL: 'https://dream-gateway.livepeer.cloud',
});

const result = await livepeer.generate.textToImage({
  modelId: 'ByteDance/SDXL-Lightning',
  prompt: 'A coffee shop at sunset, photorealistic',
  width: 1024,
  height: 1024,
});

console.log(result.imageResponse?.images[0]?.url);

The SDK wraps the same REST surface and handles request shaping, error types, and retries.

livepeer-ai-python

Install the package:

pip install livepeer-ai

Call the gateway:

from livepeer_ai import Livepeer

client = Livepeer(server_url="https://dream-gateway.livepeer.cloud")

result = client.generate.text_to_image(request={
    "model_id": "ByteDance/SDXL-Lightning",
    "prompt": "A coffee shop at sunset, photorealistic",
    "width": 1024,
    "height": 1024,
})

print(result.image_response.images[0].url)

Speakeasy-generated from the same OpenAPI spec as the TypeScript SDK.

Pipeline Models

The text-to-image pipeline accepts any diffusion model supported by ai-runner. Two models are kept warm on the network and respond immediately:

Model	Notes
ByteDance/SDXL-Lightning	Fast, four-step inference. Recommended default
SG161222/RealVisXL_V4.0_Lightning	Photorealistic, four-step variant

Other supported models (cold-start on first request, then warm):

• stabilityai/stable-diffusion-xl-base-1.0
• SG161222/Realistic_Vision_V6.0_B1_noVAE
• runwayml/stable-diffusion-v1-5
• prompthero/openjourney-v4
• stabilityai/sd-turbo, stabilityai/sdxl-turbo (limited-commercial licence)

A cold model takes between 30 seconds and a few minutes for the first response while an orchestrator loads it into GPU memory.

For all available pipelines (image-to-image, image-to-video, audio-to-text, LLM, and more), see .

Common Errors

No orchestrator available for model

No orchestrator on the network currently has capacity for the requested model. Retry after a short wait, or pick one of the warm models above. The pipeline reference page lists which models are kept warm.

Cold model load (slow first request)

The first request to a non-warm model triggers an orchestrator to load weights into GPU memory. Subsequent requests against the same model are fast. The community gateway does not stream progress; the request blocks until the model is ready.

Validation error on `width` or `height`

Most SDXL-family models accept multiples of 64 between 512 and 1024. Drop down to 768 or 1024 squared if a request fails on dimensions.

Rate limit on the community gateway

The community gateway is shared infrastructure. High-volume callers should move to a paid gateway or self-host. A paid gateway accepts the same request shape; pass a Bearer token in the Authorization header.

The community gateway is shaped for experimentation. Production workloads should run against a paid gateway or a self-hosted gateway. See .

You now have a working inference call against the Livepeer network. The AI pipelines page documents all nine batch endpoints with request schemas and curl examples for each.

Next Steps

AI Pipelines

The eleven native pipelines, request shapes, and what each one solves.

AI SDKs Overview

Full SDK surface: error types, retries, streaming, batching.

Model Support

Warm models per pipeline, VRAM requirements, custom-model paths.

AI Image Generation App

Wrap this quickstart in a working Next.js app.