Image generation

greenrouter exposes image models (gpt-image-2, gemini-3.1-flash-image) through OpenAI-compatible image endpoints: /v1/images/generations (text → image) and /v1/images/edits (image-to-image with reference images). Both are live and authenticate with an sk-ru-… key, exactly like chat.

Endpoints

POST https://api.ru-llm.relay2.xyz/v1/images/generations
POST https://api.ru-llm.relay2.xyz/v1/images/edits

Generate an image

curl https://api.ru-llm.relay2.xyz/v1/images/generations \
  -H "Authorization: Bearer sk-ru-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "A red panda coding at a laptop, soft studio lighting",
    "n": 1,
    "size": "1024x1024",
    "quality": "high"
  }'

Field	Type	Required	Description
model	string	Yes	Image model ID — gpt-image-2 or gemini-3.1-flash-image.
prompt	string	Yes	Text description of the image to generate.
n	integer	No	Number of images to generate. Default: 1.
size	string	No	Output resolution. gpt-image-2 accepts 1024x1024, 1536x1024, 1024x1536, and auto.
quality	string	No	Rendering quality for gpt-image-2: low, medium, high, or auto.

Note

The gateway routes on model and forwards the rest of the body to the provider verbatim, so any field the underlying image model accepts is passed straight through. The fuller settings passthrough is tracked in RUL-103 — this page is the home for those knobs as they land.

Response

The response follows the OpenAI images envelope. gpt-image-2 always returns base64-encoded image data (b64_json); the gateway relays the payload unchanged.

{
  "created": 1710000000,
  "data": [
    {"b64_json": "iVBORw0KGgoAAAANSUhEUgAA..."}
  ],
  "usage": { "input_tokens": 42, "output_tokens": 1568 }
}

Decode the first image to a PNG file:

curl ... | python3 -c "import sys,json,base64; \
open('out.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"

Image-to-image (edits)

To transform or extend an existing image, send a multipart/form-data request to /v1/images/edits with one or more reference images and a prompt. This is the surface for image-to-image and inpainting.

curl https://api.ru-llm.relay2.xyz/v1/images/edits \
  -H "Authorization: Bearer sk-ru-YOUR_KEY" \
  -F model="gpt-image-2" \
  -F prompt="Put the panda on a snowy mountain at dawn" \
  -F image[]="@panda.png" \
  -F size="1024x1024"

Form field	Required	Description
model	Yes	Image model ID (gpt-image-2).
prompt	Yes	What to change or generate from the reference image(s).
image[]	Yes	One or more reference image files (the source to edit).
mask	No	Optional PNG mask marking the editable region (inpainting).
n, size, quality	No	Same meaning as in generations.

The response is the same OpenAI images envelope (base64 b64_json) as generations.

Note

Image variations (/v1/images/variations) are not offered — the upstream provider does not support them for these models. Use /v1/images/edits with a reference image instead.

Billing

Billing basis depends on the model — see the Models & Pricing page and the console Pricing page for live rates.

• gpt-image-2 is billed per token, on the input_tokens / output_tokens the response usage block reports. Those counts scale with size and quality, so token billing is naturally resolution- and quality-aware.
• gemini-3.1-flash-image is billed per generated image (the per_image rate), charged once per image in the data array.

The charge settles against a fail-closed funds hold, so a zero-balance account is rejected with 402 before the request reaches the provider.

Troubleshooting

Symptom	Cause	Fix
401 Unauthorized	Key missing or wrong	Pass a valid sk-ru-… key.
402 Payment Required	Zero balance	Top up in the console → Billing.
400 not an image model	Model ID is a chat/embedding model	Use gpt-image-2 or gemini-3.1-flash-image.
400 unknown model	Model ID not in the catalog	Check GET /v1/models.
400 expected multipart	/v1/images/edits sent as JSON	Send edits as multipart/form-data (use -F, not -d).