Gemini Native Image Generation

Use the official Gemini generateContent REST shape for native image generation and image editing models.

Gemini native image generation uses Google's official generateContent REST request shape: the model is carried in the path, text and images are sent through contents[].parts[], and image output is requested with generationConfig.responseModalities.

If your client already uses the Gemini SDK or Google REST format, you can point it at the Models Hub Gemini-compatible endpoint without converting the request body to the OpenAI Images format.

Endpoints

Operation	Endpoint
Gemini native content generation	`POST /v1beta/models/{model}:generateContent`
Gemini native content generation	`POST /v1/models/{model}:generateContent`

{model} is the Gemini image model ID, such as gemini-3.1-flash-image-preview or another Gemini image model available to your account.

Request Shape

Field	Type	Required	Description
`contents`	array	Yes	Request content. A single-turn request usually has one `role: "user"` item.
`contents[].role`	string	No	`user` or `model`. Use `user` for a single-turn request.
`contents[].parts`	array	Yes	Multimodal parts. Each part should carry one content type.
`parts[].text`	string	No	Text prompt.
`parts[].inlineData`	object	No	Base64 inline media using the official camelCase fields.
`inlineData.mimeType`	string	Yes	Input image MIME type, such as `image/png`, `image/jpeg`, or `image/webp`.
`inlineData.data`	string	Yes	Base64 image data without a `data:image/...;base64,` prefix.
`generationConfig.responseModalities`	string array	Yes	Image generation must include `IMAGE`. Use `["TEXT", "IMAGE"]` if you also want text.
`generationConfig.imageConfig.aspectRatio`	string	No	Output aspect ratio, such as `1:1`, `4:3`, `16:9`, `9:16`, or `21:9`.
`generationConfig.imageConfig.imageSize`	string	No	Output image tier, such as `512`, `1K`, `2K`, or `4K`, depending on model support.
`systemInstruction`	object	No	Gemini official system instruction shape.
`safetySettings`	array	No	Gemini official safety settings.
`tools`	array	No	Gemini official tools.

Text To Image

curl -X POST "https://modelsok.com/v1beta/models/gemini-3.1-flash-image-preview:generateContent" \
  -H "Authorization: Bearer $MODELSOK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [
      {
        "role": "user",
        "parts": [
          {
            "text": "Create a polished product photo with clean composition, soft natural light, and crisp detail."
          }
        ]
      }
    ],
    "generationConfig": {
      "responseModalities": ["IMAGE"],
      "imageConfig": {
        "aspectRatio": "4:3",
        "imageSize": "2K"
      }
    }
  }'

Reference Image Editing

When sending an image through inlineData, data must be the complete Base64 string. Do not include a Data URL prefix and do not send a truncated log capture.

{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "text": "Remove all people from the scene while preserving the room structure. Make the final image clean, airy, and cinematic."
        },
        {
          "inlineData": {
            "mimeType": "image/png",
            "data": "BASE64_IMAGE_DATA"
          }
        }
      ]
    }
  ],
  "generationConfig": {
    "responseModalities": ["IMAGE"],
    "imageConfig": {
      "aspectRatio": "16:9",
      "imageSize": "2K"
    }
  }
}

Response

The response keeps the native Gemini structure. Generated image data is usually returned at:

candidates[].content.parts[].inlineData.data

The MIME type is returned at:

candidates[].content.parts[].inlineData.mimeType

Common Errors

Problem	Description
Missing `responseModalities`	Image generation must request `IMAGE`; an empty list is equivalent to text-only output.
Truncated Base64	`inlineData.data` must be complete. Truncated payloads commonly produce `400 invalid argument`.
Data URL prefix included	`inlineData.data` expects raw Base64, not `data:image/png;base64,...`.
Mixed new-interface fields	This endpoint uses `contents`, `inlineData`, and `generationConfig`. Do not send `interactions` fields such as `input`, `type: "image"`, or `response_format` in a `generateContent` body.
`imageConfig` sent to a non-image model	`imageConfig` is only valid for models that support image generation configuration.

Difference From Generic Images

/v1/images/generations is the OpenAI Images-compatible entrypoint for unified image model calls. Gemini native image generation preserves Google's official generateContent request and response shape, which is best for clients already built around the Gemini protocol.