Choose AI Image Editing Models by Inputs and Controls

The image-edit endpoint accepts up to ten references and can return one to four images. Custom width and height range from 256 to 4096 pixels, with an aspect-ratio constraint.

Custom dimensions override preset size controls. Charges depend on output size and image count.

Use up to ten references, with 1K or 2K output and PNG, WebP or JPEG formats. The guide recommends no more than four references when fidelity is the priority.

A mask cannot be combined with transparent output. Transparent backgrounds require PNG or WebP.

Supply up to ten JPEG, PNG or WebP images. Basic quality produces 1K output and high produces 2K. Prompts accept 3–5000 characters.

Each reference is limited to 30 MB and 4.19 million pixels. Account access requirements apply.

Basic, high and ultra map to 2K, 3K and 4K. Supply JPEG, PNG or WebP files up to 30 MB each.

Prompts must be 3–3000 characters. Choose an explicit supported aspect ratio; auto is not listed.

Explain each reference’s role in the order supplied. JPEG, PNG and WebP inputs are supported, up to 30 MB per file.

Higher resolutions use more credits; check labels, faces and product geometry after editing.

The SeeAPI editing endpoint accepts JPEG, PNG and WebP inputs up to 30 MB each, with 1K, 2K or 4K output.

Provider-specific fields and output costs differ; check the selected route.

Use up to eight source images and select the output aspect ratio, resolution and format.

This model has a higher access requirement in the public catalog; check account eligibility.

The default SeeAPI route accepts up to 16 references and returns one image at 1K, 2K or 4K. The fal route adds mask, quality and native multi-output controls.

Mask and multi-output settings belong to the fal route; do not send them to the default SeeAPI route.

The default SeeAPI route supports up to 16 source images and one 1K–4K result. Select fal for its documented mask, quality, format and multi-output controls.

Provider inputs and billing differ. Choose the provider before selecting editing parameters.

Supply up to ten reachable JPEG, PNG or WebP URLs, with a maximum of 30 MB per source image.

Do not carry resolution settings over from other Nano Banana variants; follow this endpoint's schema.

The editing endpoint accepts up to ten JPEG, PNG or WebP image URLs, up to 30 MB each, plus a text instruction.

Compare against newer variants with the same source images; model-family membership does not imply identical controls.

Use up to nine references and up to two bounding boxes per source image. The editing API offers 1K or 2K resolution, sequential generation and palette controls.

Bounding boxes must align with the input array. Color palettes are not applied when sequential generation is enabled.

Choose automatic, opaque or transparent backgrounds. On Replicate, transparent output requires PNG or WebP.

SeeAPI background controls require 1K. Resolution and aspect-ratio combinations vary by provider.

The current input schema accepts up to four source images. Output aspect ratios are 1:1, 3:2 and 2:3.

Use the schema maximum of four inputs; do not assume the higher reference limits of GPT Image 2 or 2.5.

Use up to ten JPEG, PNG or WebP reference images, up to 30 MB each. Select 1K, 2K or 4K output and an optional seed.

Supported aspect ratios and controls differ from Seedream 5.0 variants.

Describe the change and the details to preserve. Compare prompt upsampling on and off using the same source.

Auto aspect ratio is documented for Replicate; SeeAPI uses explicit ratios.

Supply exactly one reference image and an editing instruction of up to 2,000 characters. Review the selected provider's aspect-ratio and format options.

When automatic translation is disabled, the SeeAPI guide requires an English prompt.

The SeeAPI provider offers negative prompts, 2–49 inference steps, guidance scale 1–20 and a seed.

Controls vary by provider. A fixed seed does not guarantee identical results.

The image-to-image endpoint accepts exactly one JPEG, PNG or WebP image up to 10 MB and returns one image per request.

This editing endpoint does not accept the multiple-reference inputs supported by the models above.