Image Generation

Google Gemini image generation models are exposed through /v1/images/generations for text-to-image, image-to-image, and multi-reference image generation. Imagen models run as text-to-image through this gateway. For OpenAI gpt-image-2 text-to-image usage, see Gateway image generation.

Text-to-image

curl "$LLMTR_BASE_URL/v1/images/generations" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-3.1-flash-image",
    "prompt": "A minimalist modern workspace with warm light",
    "size": "1K",
    "aspect_ratio": "16:9",
    "n": 1
  }'

Image-to-image (reference image)

Upload a source image via Files API first, then pass the file_id. This singular field remains supported for backward compatibility:

curl "$LLMTR_BASE_URL/v1/images/generations" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-3.1-flash-image",
    "prompt": "Turn this product photo into an ad poster",
    "file_id": "file_xxx",
    "size": "1K",
    "aspect_ratio": "1:1",
    "n": 1
  }'

Multiple reference images

Use file_ids to send more than one reference image to Gemini image models. The total reference image limit is 8 across file_id, file_ids, image_url, and image_urls.

curl "$LLMTR_BASE_URL/v1/images/generations" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-3.1-flash-image",
    "prompt": "Combine these two product photos into one ad composition",
    "file_ids": ["file_front", "file_back"],
    "size": "1K",
    "aspect_ratio": "16:9",
    "n": 1
  }'

For remote or inline image references, image_url and image_urls are also supported. File-based references are recommended for larger images and references reused across requests.

file_ids vs image_urls — when to use which

Dimension	file_ids (upload via /v1/files first)	image_urls (data URL or https URL)
HTTP round trips	Two (upload + generations)	One
Large images (> 1–2 MB)	Preferred — JSON body stays small	Base64 in JSON inflates by ~33%
Reusing the same reference across requests	Preferred — upload once, reuse file_id for 7 days	Resent on every request
One-off, fast experiments	Slower (two requests)	Preferred — pass the data URL or https URL directly
Referencing a public hosted image	Requires download + upload	Pass the https URL directly
Cleanup	Files are deleted automatically after 7 days	No storage; per-request only

Rule of thumb: for try-on, image-to-image and repeated compositions with large references use file_ids; for one-off compositions or direct public-URL references use image_urls. A single request can combine both; the total reference image limit is 8.

Try-on: merging two images

To dress a person from a model photo with a separate garment photo, compose a product into a scene, or otherwise merge two reference images, pass both file_ids in a single /v1/images/generations request. The flow is three steps: upload each image through /v1/files, collect the returned file_ids, and send them as a file_ids array.

# 1) Upload the model photo, capture the file_id
curl "$LLMTR_BASE_URL/v1/files" \
  -H "Authorization: Bearer llmtr-your_key" \
  -F "provider=google" \
  -F "purpose=vision_reference" \
  -F "file=@./model.jpg"
# → {"id": "file_model_xxx", ...}

# 2) Upload the garment photo, capture the file_id
curl "$LLMTR_BASE_URL/v1/files" \
  -H "Authorization: Bearer llmtr-your_key" \
  -F "provider=google" \
  -F "purpose=vision_reference" \
  -F "file=@./garment.jpg"
# → {"id": "file_garment_yyy", ...}

# 3) Merge both references in one request
curl "$LLMTR_BASE_URL/v1/images/generations" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-3.1-flash-image",
    "prompt": "Dress the person in the first image with the t-shirt from the second image. Keep the lighting and shadows consistent with the model photo, and preserve the pattern and colors from the garment photo.",
    "file_ids": ["file_model_xxx", "file_garment_yyy"],
    "n": 1
  }'

All reference-image flows, including try-on, run through /v1/images/generations. On LLMTR, /v1/images/edits is open only for xAI image models; calling it with a Google model returns a 400 whose response body points back to this flow. Imagen models support text-to-image only and cannot be called with file_ids.

Parameters

Field	Values
size	512, 1K, 2K
aspect_ratio	1:1, 3:4, 4:3, 9:16, 16:9
n	1-4
file_id	Optional single reference image
file_ids	Optional list of reference images
image_url	Optional single remote or inline reference image
image_urls	Optional list of remote or inline reference images

When file_id, file_ids, image_url, and image_urls are combined, duplicate values are deduplicated. Google Imagen models do not support reference image fields on this gateway; send prompt-only text-to-image requests.

Response

{
  "created": 1739200000,
  "data": [
    {
      "b64_json": "iVBORw0KGgoAAAANSUhEUgAA...",
      "revised_prompt": "A minimalist modern..."
    }
  ]
}

Pricing

Billed per image by native unit. See the model card in Dashboard > Models.