Free image API

Quick start

curl https://seinari.ai/v1/generate \
  -H 'Content-Type: application/json' \
  -d '{"prompt": "a lighthouse on a cliff at sunset, oil painting", "rewrite": false}'

The answer is JSON with the picture in it, base64-encoded:

{"model":"fast","rewrite":false,"seed":1180428871,"size":512,"type":"image/jpeg",
 "gpu_ms":465.2,"ms":688.0,"image":"/9j/4AAQSkZJRgABAQ…"}

Or get the picture itself:

curl https://seinari.ai/v1/generate -H 'Accept: image/jpeg' \
  -d '{"prompt": "a red fox in fresh snow", "seed": 7}' -o fox.jpg

A plain link works too, so you can put it straight in an <img> tag:

https://seinari.ai/v1/generate?prompt=a+red+fox+in+fresh+snow&seed=7

Models

Model	Underlying model	Size
`fast`	FLUX.2 Klein 4B	512×512
`quality`	FLUX 3 Klein (Klein 3)	1024×1024

POST /v1/generate

Field	Meaning
`prompt`	What to draw, 1 to 400 characters. Required.
`seed`	A whole number from 0 to 2147483647. Repeats the drawing request, but different cards or runtime updates can change the picture. Leave it out for a new random seed.
`model`	Which model draws it: `fast` (the default) or `quality`. /v1/models lists them.
`size`	Optional; each model draws one size (512 for `fast`, 1024 for `quality`).
`rewrite`	Quality always expands the prompt before drawing; it has no rewrite opt-out. Fast draws your original prompt. /v1/models identifies rewritten models.
`format`	`"json"` (the default for POST) or `"image"` (the default for GET, and for requests that send `Accept: image/jpeg`).

GET /v1/generate?prompt=…&seed=…&model=… takes the same fields as a query string.

Picture answers carry the same facts as headers: X-Seed, X-Model, X-Rewrite and X-GPU-Ms. CORS is open, so it works from any web page.

Nothing is kept

Every picture is drawn for your request, sent to you, and forgotten. Our service does not save prompts or pictures; answers are marked Cache-Control: no-store. If you want to keep a picture, save it when it arrives.

Errors

Code	Meaning
`400`	Something in the request is wrong; `error` says what.
`429`	Shared capacity is busy: the fleet rejects new work when expected queue wait exceeds 4 seconds or 512 jobs are waiting. The API also caps its own in-flight requests at 256. These are capacity limits, not per-user quotas. Wait `Retry-After` seconds before retrying.
`503`	No serving capacity is available, or Quality or the fallback is not ready. Wait `Retry-After` seconds before retrying.

At 116 images per second, Fast measured 415 ms client p50 from Helsinki to the North American fleet; router p50 was 136 ms.

gpu_ms is the worker generation wall time in milliseconds, including text encoding and CPU work; it is not a GPU-only measurement.

Pictures drawn through the API count toward the live counter on seinari.ai. The counter keeps only the number of pictures and their worker generation time.