Image generation

POST /v1/images/generations turns a text prompt into one or more images, OpenAI-style. It works with any image model — sdxl, flux-schnell, qwen-image.

Images are returned as base64 PNGs (b64_json). Hosted url responses are not supported yet.

Generate an image

from openai import OpenAI
import base64
 
client = OpenAI(api_key="sk-crustoff-...", base_url="https://api.crustoff.app/v1")
 
img = client.images.generate(
    model="sdxl",
    prompt="a red fox in the snow, golden hour, highly detailed",
    size="1024x1024",
)
 
with open("fox.png", "wb") as f:
    f.write(base64.b64decode(img.data[0].b64_json))

import OpenAI from "openai";
import fs from "node:fs";
 
const client = new OpenAI({
  apiKey: process.env.CRUSTOFF_API_KEY,
  baseURL: "https://api.crustoff.app/v1",
});
 
const img = await client.images.generate({
  model: "sdxl",
  prompt: "a red fox in the snow, golden hour, highly detailed",
  size: "1024x1024",
});
 
fs.writeFileSync("fox.png", Buffer.from(img.data[0].b64_json, "base64"));

curl https://api.crustoff.app/v1/images/generations \
  -H "Authorization: Bearer $CRUSTOFF_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sdxl",
    "prompt": "a red fox in the snow, golden hour, highly detailed",
    "size": "1024x1024"
  }'

Request parameters

Field	Type	Default	Notes
`model`	string	—	An `image` model id
`prompt`	string	—	Required, non-empty
`n`	int	`1`	Number of images, `1`–`4`
`size`	string	`"1024x1024"`	`WxH`; each side `256`–`1536`, a multiple of `8`
`response_format`	string	`"b64_json"`	Only `b64_json` is supported
`negative_prompt`	string	—	What to avoid (ignored by FLUX, which is guidance-distilled)
`seed`	int	random	Set for reproducible output
`steps`	int	per-model	Denoising steps. Clamped to `1`–`60`; defaults to 30 (SDXL) / 4 (FLUX schnell)

Response

{
  "created": 1718900000,
  "data": [
    { "b64_json": "iVBORw0KGgoAAAANSUhEUgAA..." }
  ]
}

Each data[i].b64_json is a base64-encoded PNG. The response header x-prox-request-id correlates the request with your usage and billing.

Choosing a model

Model	Speed	Notes
`sdxl`	Fast	Great all-rounder, broad style range
`flux-schnell`	Fastest	Modern 12B quality in 1–4 steps
`qwen-image`	Slower	Flagship 20B quality, excellent text rendering

Billing

Images are billed per megapixel of output — the actual pixels, with no rounding 1024×1024 up to 2 MP like some providers. The cost is fully known before generation:

megapixels = width × height × n ÷ 1,000,000
cost       = megapixels × model_rate

A single 1024x1024 image is 1.05 MP. On sdxl ($0.0015/MP) that’s about $0.0016. See Pricing & billing for all rates.

⚠️

If generation fails on our side, you are not charged — the credit hold is released in full. A cold model may take longer on the first request while a GPU spins up.

Text generation Pricing & billing