Core concepts

Image generation

From zero: install openai, set base_url, call images.generate. Pass optional reference images to edit. Billed per image or per token in RODI.

POSThttps://api.rodiumai.io/v1/images/generations

RodiumAi exposes OpenAI-compatible image generation. Text-to-image works with OpenAI GPT Image and Google Gemini Image models. Pass image or images on the same endpoint to edit: GPT Image requests are routed to the provider's edit API, Gemini Image models edit natively.

Model ids & input images

Use provider-scoped ids such as openai/gpt-image-1.5 or google/gemini-3.1-flash-image. Input images (data URL or b64_json) work on GPT Image and Gemini Image models; gs:// URIs work on Gemini Image only. Responses contain b64_json. List models with GET /v1/models.

When to use

Product heroes, ads, and social creatives from a text brief.
Edit or restyle a still with GPT Image or Gemini Image (image / images).
Batch variants (n, size, quality) or feed a still into Veo.

Recipes

Product hero

Describe lighting, angle, and brand mood in the prompt. Start with 1024x1024 medium quality.

Note: Billed per image in RODI, check GET /v1/pricing before large n.

Image edit

Use a GPT Image or Gemini Image model and pass image (or images[]) as { b64_json } or a data URL. HTTP(S) URLs are rejected.

Note: gs:// references are accepted by Gemini Image models only; send b64_json or a data URL to GPT Image.

Image → video

Generate or edit a still here, then POST /v1/videos/generations with image.b64_json (not an HTTPS URL).

Note: See the video guide image-to-video recipe.

Text → image examples

…

Image → image (edit)

Send prompt plus image or images (max 14). Same encodings as video: { b64_json }, data URL, or gs://. HTTP(S) image URLs are rejected.

…

Request parameters

Parameter	Type	Required	Description
model	string	Required	Image model id (e.g. openai/gpt-image-1.5, google/gemini-3.1-flash-image). List them with GET /v1/models.
prompt	string	Required	Text description of the image to generate or edit instruction.
image	object \| string	Optional	Optional reference image to edit. Accepts { b64_json }, a data URL or raw base64; gs:// URIs on Gemini Image models only. HTTP(S) URLs are rejected. On GPT Image models the request is sent to the provider's edit API. Aliases: input_image, image_url.
images	array	Optional	Several reference images (same encodings as image), max 14. Alias: input_images. Use for multi-image edits and merges.
n	integer	Optional	Number of images (1–10). Default 1.
size	string	Optional	Output size, e.g. "1024x1024", "1536x1024", "1024x1536" (mapped to an aspect ratio on Gemini Image models).
quality	string	Optional	Quality when supported (GPT Image: low, medium, high). Also selects the price tier of flat-priced models.
mask	object \| string	Optional	GPT Image edits only: PNG mask (same encodings as image) marking the area to change.
background / output_format	string	Optional	Forwarded to GPT Image models (e.g. background: transparent, output_format: webp).

API reference: images

Models and input images

Family	Text to image	Input images	Encodings
GPT Image (`openai/gpt-image-1.5`, `openai/gpt-image-1-mini`, …)	Yes	Yes: the gateway calls the provider's edit API; `mask` supported	`{ b64_json }`, data URL, raw base64
Gemini Image (`google/gemini-3.1-flash-image`, `google/gemini-3-pro-image`, …)	Yes	Yes, up to 14 (`image` or `images`)	`{ b64_json }`, data URL, raw base64, `gs://`

Remote http(s):// image URLs are never fetched by the gateway: encode the file as base64. Keep the whole request under 10 MiB, or it is rejected with 413.

Response

Images come back as base64 in data[].b64_json (OpenAI routes also add mime_type). GPT Image models always answer in base64, so response_format is not needed.
Decode b64_json and write the bytes to a file, as in the examples above. Some models may return data[].url instead; handle both.
Billing: per image (flat-priced models, by quality and size) or per token (GPT Image), see Pricing.