Image Generation APIs: GPT Image 2 & Gemini

Generate images with GPT Image 2 through the OpenAI-compatible Images API or with Nano Banana Pro and Nano Banana 2 through the native Gemini API.

This page keeps two separate image-generation API contracts in one place. They share your Modelflare API key and Credits, but their request and response formats are not interchangeable.

gpt-image-2 uses the OpenAI-compatible /v1/images/generations and /v1/images/edits contracts.
gemini-3-pro-image and gemini-3.1-flash-image use the native Gemini generateContent contract.

Choose the model family before copying an example. Do not send Gemini fields to the GPT endpoint or GPT Images fields to a Gemini endpoint.

1. GPT Image 2 — OpenAI-compatible Images API

Generate PNG images or edit reference images supplied by public URL or local upload with gpt-image-2 through the OpenAI-compatible Images API. Each request returns one image and supports medium or high quality, signed URL or base64 output, and optional SSE.

Choose an endpoint and mode

Use https://origin.modelflare.dev/v1 for ordinary synchronous JSON requests that may take several minutes.
Use https://modelflare.dev/v1 with stream=true for long requests through Cloudflare. Modelflare sends transport-only SSE heartbeats until the final image is ready.
Do not redirect an in-flight POST between hosts. Choose the base URL before submitting the request.

Synchronous request

curl https://origin.modelflare.dev/v1/images/generations \
  -H "Authorization: Bearer YOUR_MODELFLARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "A quiet lakeside cabin at dawn, cinematic natural light",
    "n": 1,
    "quality": "medium",
    "size": "1024x1024",
    "response_format": "url",
    "output_format": "png"
  }'

A URL response contains a private Modelflare link that expires after one hour:

{"created":1710000000,"data":[{"url":"https://modelflare.dev/v1/images/assets/.../content?expires=...&signature=..."}]}

Set response_format to b64_json to receive the validated PNG inline instead of storing an API output object.

The live Models & Pricing page shows medium and high price points for common sizes. Billing uses the exact requested quality and dimensions, ignores ordinary routing-group multipliers, and records the final charge in usage logs.

Edit a reference image

Send one or more publicly reachable HTTP(S) image URLs to POST /v1/images/edits. Each images item contains exactly one image_url:

curl https://origin.modelflare.dev/v1/images/edits \
  -H "Authorization: Bearer YOUR_MODELFLARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "Replace the background with a clean white photography studio",
    "images": [
      {"image_url": "https://example.com/input.png"}
    ],
    "size": "2048x2048",
    "response_format": "url"
  }'

Or upload one local PNG, JPEG, or WebP file with multipart/form-data:

curl https://origin.modelflare.dev/v1/images/edits \
  -H "Authorization: Bearer YOUR_MODELFLARE_API_KEY" \
  -F "model=gpt-image-2" \
  -F "prompt=Replace the background with a clean white photography studio" \
  -F "image=@/path/to/input.png" \
  -F "size=2048x2048" \
  -F "response_format=url"

The multipart request must contain exactly one image file and stay within 20 MiB in total. The reference URL or uploaded file is sent to the selected upstream provider; URL inputs must remain reachable while fetched. Modelflare does not retain either reference input as an API output asset.

Streaming request

curl -N https://modelflare.dev/v1/images/generations \
  -H "Authorization: Bearer YOUR_MODELFLARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-image-2","prompt":"A white cat running through a neon street in the rain","n":1,"quality":"high","size":"1024x1024","response_format":"url","stream":true}'

SSE comments are heartbeats, not model output. Stop after image_generation.completed and data: [DONE].

Supported fields

model: required; exactly gpt-image-2.
prompt: required; up to 32,000 Unicode characters.
n: optional; currently must be 1.
quality: medium (default) or high.
size: WIDTHxHEIGHT, using multiples of 16, a maximum edge of 3840 px, an aspect ratio up to 3:1, and 0.66–8.29 MP.
response_format: url (default) or b64_json.
output_format: currently png only.
stream: optional boolean.
images: required only for /v1/images/edits; a non-empty array of objects containing one public HTTP(S) image_url each. URLs with embedded usernames or passwords are rejected.
image: the alternative multipart input for /v1/images/edits; exactly one valid PNG, JPEG, or WebP file.

Generation requests reject reference images. The edit endpoint accepts either JSON URL references or one multipart local upload; multiple uploaded files, data/base64 inputs, masks, input_fidelity, variations, partial images, and unknown fields are rejected. Failed upstream, validation, download, or storage operations are not delivered as successful images and reserved Credits are returned through the request-id-safe refund path.

Use a normal API key. Channel-pinned keys are rejected, and a key with model limits enabled must include gpt-image-2.

2. Gemini Image — native Gemini API

Use Nano Banana Pro or Nano Banana 2 through the native Gemini request and response format. These models do not use /v1/images/generations.

Endpoint and long-request mode

Synchronous: POST https://modelflare.dev/v1beta/models/{model}:generateContent.
Recommended for image generation: POST https://modelflare.dev/v1beta/models/{model}:streamGenerateContent?alt=sse.
origin.modelflare.dev does not expose /v1beta. Use the canonical domain for Gemini native requests, and prefer streaming for long jobs.

Streaming request

This example uses Nano Banana 2. Replace the model in the URL with gemini-3-pro-image to use Nano Banana Pro.

curl -N "https://modelflare.dev/v1beta/models/gemini-3.1-flash-image:streamGenerateContent?alt=sse" \
  -H "x-goog-api-key: YOUR_MODELFLARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{
      "role": "user",
      "parts": [{"text": "Create a clean blue circle centered on a white background. No text."}]
    }],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"],
      "imageConfig": {"aspectRatio": "1:1", "imageSize": "1K"}
    }
  }'

Response contract

Each SSE data: event is a native Gemini JSON response. The final image is base64 in candidates[].content.parts[].inlineData.data, with its MIME type in inlineData.mimeType. Text parts may appear in the same response. Finish when the final candidate reports finishReason=STOP or the stream closes; GPT events such as image_generation.completed are not used here.

Models, resolutions, and estimated prices

Model	Resolution	Estimated price per image
Nano Banana Pro	1K / 2K	US$0.06720
Nano Banana Pro	4K	US$0.12000
Nano Banana 2	0.5K	US$0.02241
Nano Banana 2	1K	US$0.03360
Nano Banana 2	2K	US$0.05040
Nano Banana 2	4K	US$0.07560

Use uppercase K in imageSize. Prices are estimates for one generated image; the final charge may vary slightly with request content and is recorded in usage logs.

Use a normal Modelflare API key through x-goog-api-key. The key must be able to use the gemini-image group, and model-limited keys must include the selected Gemini image model.