Image generation API models on Sume: edits, references, and masks

Sume's image generation API, POST /v1/images, serves 17 catalog models through one request shape. 13 of them edit from reference images, 4 are text-to-image only, and just the two ChatGPT Image 2.5 ids accept a mask. GET /v1/images/models reports what each one accepts.

The table and rules come from the Image API docs and the capability descriptors in that catalog, read 2026-09-26. sume/auto has no catalog row: Sume picks the family and never discloses which one ran.

How can I tell whether a model edits images?

Check two fields on its catalog row. An edit-capable model reports architecture.input_modalities as ["text", "image"] and an input_references range whose maximum is above zero. A text-to-image-only model reports ["text"] and {"min": 0, "max": 0}; send it a reference and the call fails with 400 unsupported_parameter.

The per-model record at GET /v1/images/models/{model_id}/endpoints repeats the same supported_parameters, which the docs call the definitive parameter set for that endpoint.

Which image models can edit from reference images?

The two ChatGPT Image 2.5 ids take up to 16 references, and every other edit-capable id takes up to 10. The text-only ids are qwen/qwen-image-max, google/imagen-4-fast, google/imagen-4-ultra, and recraft/recraft-v4. Each row below is what the catalog reports for that id:

How many reference images can I send, and in what form?

Up to the model's input_references maximum. Each entry is an object with type: "image_url" and an image_url.url, and every URL must be public HTTPS: localhost, private-network, and non-HTTPS URLs are rejected before submission. A request with more references than the model allows is refused with a 400 error. Image generation with reference images shows the full edit call.

One family behaves differently once references arrive. Seedream 4.5 lists 10, but a Seedream 4.5 request that includes them returns a single image and does not apply output_format; the Seedream 4.5 API has the details.

Which models support masked edits?

Only openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst. Their descriptors list mask_url, an optional public HTTPS mask for ChatGPT Image 2.5 edits, and background, whose values are auto, transparent, and opaque. No other row lists either field, so sending one to another model returns 400 unsupported_parameter. The GPT Image 2.5 API walks through a masked edit.

Where are sizes, quality, and image counts covered?

In their own guides, because they vary along different lines: aspect ratios and custom sizes covers aspect_ratio and image_size, and 4K, quality, and image count covers resolution, quality, and n. Bare legacy ids such as nano-banana-2 still work as aliases for the org/slug form.

How are image models priced?

Per output image. Each endpoints record carries a pricing line with billable: "output_image" and unit: "image", and that line already includes Sume's margin, plus a 5.5% agent fee by default. On the ChatGPT Image ids the line is one reference figure for high quality at 1024×1024, and each request is estimated from its size and quality.

Failed or cancelled generations are not billed. API pricing lists the Image Router at provider list rates times Sume's margin, with per-model rates in the catalog.