Image generation

Generate images with a model carrying the image category in the catalog.

cURL

curl https://api.resetdata.ai/api/v1/images/generations \
  -H "Authorization: Bearer $RESETDATA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "black-forest-labs/flux-1-schnell",
    "prompt": "A harbour at dawn, long exposure, muted colours."
  }'

Python

resp = client.images.generate(
    model="black-forest-labs/flux-1-schnell",
    prompt="A harbour at dawn, long exposure, muted colours.",
)

JavaScript

const resp = await client.images.generate({
  model: "black-forest-labs/flux-1-schnell",
  prompt: "A harbour at dawn, long exposure, muted colours.",
});

Parameters vary a lot here

Image models diverge more than text models. Depending on the model you may have width, height, steps, cfg_scale, seed, samples or a reference image — and the useful ranges differ too.

Check supported_parameters on the model before tuning:

curl "https://api.resetdata.ai/api/v1/models/detail?slug=black-forest-labs/flux-1-schnell" \
  -H "Authorization: Bearer $RESETDATA_API_KEY" | jq '.supported_parameters'

As everywhere, a parameter the model doesn’t declare is dropped silently — so an unsupported steps will not error, it will simply not apply. See Model parameters.

Two that behave predictably where supported:

• seed — fixes the noise so the same prompt and settings reproduce the same image. Essential for iterating on a prompt, since without it every change confounds prompt effects with random variation.
• steps — more denoising steps generally means more detail and more time. Distilled “schnell”-class models are tuned for very few steps; pushing steps up on those wastes time without improving output.

Cost and latency

Image generation is billed per image, not per token — see Usage and costs. Requesting multiple samples multiplies cost accordingly.

Image models typically run at low concurrency, so they are more prone to queueing under load than text models. Expect to wait at peak, set timeouts generously, and see Rate limits and queueing.

Prompting notes

• Describe subject, composition and lighting; naming a medium or lens does real work.
• Iterate one variable at a time with a fixed seed.
• Negative prompting, where supported, is more reliable for removing an artefact than asking for its absence in the main prompt.

You are responsible for the content you generate and how you use it. Check the model’s licence — linked from its page in the catalog — before commercial use.