Image generation

Image models use the OpenAI Images API surface — the same one client.images.generate speaks in every OpenAI SDK:

curl -s https://api.gopura.net/v1/images/generations \
  -H "Authorization: Bearer $GOPURA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bytedance/seedream-4.0",
    "prompt": "a majestic gopuram temple tower at golden hour",
    "size": "1024x1024"
  }' | jq -r '.data[0].b64_json' | base64 -d > out.png

Or with the SDK:

const image = await client.images.generate({
  model: 'bytedance/seedream-4.0',
  prompt: 'a majestic gopuram temple tower at golden hour',
  size: '1024x1024',
});
// image.data[0].b64_json

Request

Response

{
  "created": 1787443200,
  "data": [{ "b64_json": "…" }],
  "usage": { "images": 1, "cost": 0.0315 }
}

As everywhere on Gopura, the actual cost is in-band — image models are priced per generated image (see the pricing.image field on /v1/models), so the charge is exact and known even before generation starts.

Hosted URLs instead of base64

response_format: "url" returns hosted links instead of inline base64 — useful when a multi-image response would otherwise be megabytes of JSON (serverless callers, <img> tags, queues):

{
  "data": [{
    "url": "https://api.gopura.net/v1/artifacts/1f0c…9ab2",
    "mime_type": "image/png"
  }],
  "usage": { "images": 1, "cost": 0.0315 }
}

The links are unauthenticated capability URLs (the unguessable id is the secret) so they work directly in image tags — and they expire after ~48 hours. Download anything you want to keep. Pricing is identical to b64_json.

Codec, mime_type, and what response_format means

response_format selects the transport (base64 vs hosted URL), not the image codec — the same semantics as OpenAI's Images API. The codec is model-native: BFL models emit JPEG, others PNG or WebP. Every data[] entry therefore carries mime_type with the sniffed actual format — use it for file extensions and Content-Type headers instead of assuming PNG.

Latency

Generation is synchronous. The gateway enforces a 120-second upstream ceiling — a hung provider returns a clean 504 timeout and you are not billed. Observed latency from production traffic (small sample, measured 2026-08-24 — treat as indicative):

Size timeouts to the model you use; 120s covers everything we serve.

Remixing with reference images

Models whose catalog entry lists image in input_modalities accept reference images through /v1/images/edits — the same multipart surface as OpenAI's client.images.edit:

curl -s https://api.gopura.net/v1/images/edits \
  -H "Authorization: Bearer $GOPURA_API_KEY" \
  -F "model=bfl/flux-kontext-pro" \
  -F "prompt=the same scene during a monsoon thunderstorm" \
  -F "image=@reference.png" \
  | jq -r '.data[0].b64_json' | base64 -d > remixed.png

const remixed = await client.images.edit({
  model: 'bfl/flux-kontext-pro',
  prompt: 'the same scene during a monsoon thunderstorm',
  image: fs.createReadStream('reference.png'),
});

Pass several references by repeating the image field (up to 8 MB per file). Pricing is identical to generation: per output image, exact, in-band. Sending a reference to a model that doesn't accept one returns a 400 that says so.

Current image models

The openai/gpt-image-1.5 family is also available — it bills by token rather than per image, so its exact cost appears in usage.cost per request. Reference support is probe-verified per model, not assumed. Prices include margin and match the in-band usage.cost exactly. The authoritative list is always the live catalog:

Which endpoint for which model

Chat models belong to /v1/chat/completions, image models to /v1/images/generations — using the wrong endpoint returns a 400 that points you to the right one. Filter the catalog programmatically:

curl -s https://api.gopura.net/v1/models \
  | jq '.data[] | select(.architecture.output_modalities == ["image"]) | {id, pricing}'