
Learn Replicate
Open models as an API and a web playground. Every run keeps its inputs, outputs and model version, which makes it the most auditable way to generate.
What Replicate is good at
- Running image and video models without hardware
- A form per model with every input exposed
- Predictions that record inputs, outputs and version
- Pay-per-use with the API for automation
Quickstart
- Open the model pageRead the inputs, the example outputs and the version hash.
- Run in the playgroundFill the form, run, and compare outputs in the prediction view.
- Fix the seedChange one input at a time to learn the model.
- Keep the prediction linkIt is the complete record: inputs, outputs, version and time.
Settings cheat-sheet
| Setting | What it does | Start with |
|---|---|---|
| Model / version | Which model and hash | Latest, and note the hash |
| Prompt | Text input | Literal |
| Aspect ratio / size | Output dimensions | Match the delivery |
| Seed | Repeatability | Fixed while tuning |
| Guidance / steps | Where the model exposes them | Defaults |
| Output format | png, jpg, webp, mp4 | png for stills |
Replicate guides
Guides for Replicate are on their way. The quickstart and cheat-sheet above are kept current in the meantime.
The five most-asked questions
What is Replicate?
A platform that hosts open machine-learning models behind an API and a web playground. You pay per run; each model page documents its inputs and shows examples.
How do I choose a model?
Search by task (image, video, upscale, edit), check the example outputs and the version date, and read the licence on the model page.
What is a prediction?
One run of a model: its inputs, outputs, status, timing and the exact model version, kept on a page you can return to and share.
How much does a run cost?
Prices depend on the model and hardware, billed per second or per run; check the model page and the pricing page.
Can I use the outputs commercially?
The model's licence decides; Replicate links it on each model page. Record the model version with each output.