Smart Resize

The URL of the source image to resize.

The list of target dimensions to generate, each written as ``<width>x<height>`` (e.g. ``"1920x1080"``). One output image is produced per target size (multiplied by ``num_images_per_size``). The endpoint internally picks the smallest nano-banana-pro preset that fully covers each target so the final image is a pixel-perfect match with no upscaling.

Optional extra instruction that is forwarded to nano-banana-pro alongside the auto-generated resize prompt. Leave empty to let the model preserve the original image content as closely as possible.

The number of nano-banana-pro variants to produce for each target size. The response images list will contain ``len(target_sizes) * num_images_per_size`` images in total.

Hint for the minimum resolution tier to use internally. The smart-resize algorithm may pick a higher tier automatically when the requested target size is too large for this tier. This field does not affect the final output dimensions — those come from ``target_sizes``.

Options: 1K

The output image format.

Options: jpeg, png, webp

The safety tolerance level applied to the inner nano-banana-pro /edit calls. Higher numbers loosen safety filtering.

Options: 1, 2, 3, 4, 5, 6

Optional random seed forwarded to the inner nano-banana-pro /edit calls for reproducibility. The same seed is used for every target size when set.

When True, the response contains data URIs in place of CDN URLs so the client can decode the images without an extra HTTP request. Adds latency for large outputs.