replicate

Replicate API Skill

Overview

Replicate lets developers run open-source AI models in the cloud through a unified API. Run models for image generation (Flux, Stable Diffusion), video synthesis, voice, and multimodal reasoning.

Installation

npm install replicate
pip install replicate

Authentication & Setup

import Replicate from 'replicate';

const replicate = new Replicate({
  auth: process.env.REPLICATE_API_TOKEN,
});

Core API Operations

1. Run Image Generation Model (Flux.1 Schnell)

const output = await replicate.run(
  'black-forest-labs/flux-schnell',
  {
    input: {
      prompt: 'A futuristic cybernetic library with glowing neon data conduits',
      aspect_ratio: '16:9',
      num_outputs: 1,
    },
  }
);

console.log('Generated image URL:', output);

2. Asynchronous Prediction with Webhooks

const prediction = await replicate.predictions.create({
  model: 'black-forest-labs/flux-dev',
  input: {
    prompt: 'Hyper-realistic macro photography of a mechanical bumblebee',
  },
  webhook: 'https://api.myapp.com/api/webhooks/replicate',
  webhook_events_filter: ['completed'],
});

console.log('Prediction started:', prediction.id);

3. Stream Model Output

for await (const event of replicate.stream('meta/meta-llama-3-70b-instruct', {
  input: {
    prompt: 'Write a comprehensive guide to database indexing.',
  },
})) {
  process.stdout.write(event.toString());
}

AI Pitfalls & Anti-Hallucination Guidelines

• Temporary Asset URLs: Output image/video URLs returned by Replicate expire after a period of time. Always download and persist them in your own object store (S3, Cloudflare R2).
• GPU Cold Boots: Handle cold start delays gracefully in UI with progress polling.

Production Verification Checklist

• Output assets transferred to private persistent object storage
• Webhook signatures verified on completion callbacks
• Client timeouts configured to allow for GPU allocation