Replicate

Replicate is a cloud platform that makes it easy to run machine learning models via API. Instead of setting up GPU infrastructure, installing CUDA dependencies, and figuring out model weights, you call an API endpoint and get results.

## What You Can Run

Replicate hosts thousands of models across categories:

**Image Generation** - Stable Diffusion (multiple versions: 1.5, 2.1, XL, Turbo) - FLUX — the newest generation of high-quality text-to-image models - ControlNet — guided image generation with edge/depth/pose maps - Style transfer, image-to-image, inpainting, outpainting - Specialized models: anime art, photorealism, architecture, product shots

**Video Generation** - AnimateDiff — animate still images - Zeroscope — text-to-video - Various video upscaling and style transfer models

**Audio** - Whisper (OpenAI) — speech transcription - Music Gen (Meta) — music generation from text descriptions - Voice cloning and text-to-speech models

**Language Models** - Llama (Meta) — various sizes - Mistral models - Code-focused models (CodeLlama) - Specialized fine-tunes

**Computer Vision** - Object detection, segmentation, pose estimation - Image captioning and visual question answering - Background removal, super-resolution

## How It Works

Every model on Replicate gets a consistent API — call a Python or JavaScript function with a model name and input parameters, and get back a URL to the generated output. The same interface works for every model, whether image generation, video, or text.

## Deployments

For production use, you can create a Deployment — a dedicated, auto-scaling version of a model with: - Faster cold starts (warm replicas) - Predictable pricing - Custom hardware (CPU or specific GPU tiers)

## Fine-Tuning

Replicate supports training custom versions of popular models on your own data. Provide training images, specify the model to fine-tune, and get a custom model endpoint. Useful for: - Brand-consistent image generation - Custom character/style models - Domain-specific text models

## Pricing

Pay-per-second of GPU time. Rates vary by model and hardware. Most image generation runs cost a few cents. Deployments have reserved capacity pricing.