36 sneaker colorways from one photo: Nano Banana 2.1 edit, $3.60

To make 36 colorway variants of one sneaker photo on Sume, send the photo as one reference to Nano Banana 2.1, ask for n: 4 per call, and make 9 calls. At 1K that is 36 x $0.10 = $3.60. Google's pricing page describes Nano Banana 2.1 as an update to Nano Banana 2, and Sume lists it as google/nano-banana-2.1.

Why n=4 and 9 calls

The Sume catalog gives Nano Banana 2.1 an n range of 1 to 4 and an input_references range of 0 to 10. So one reference photo is well inside the limit, and 36 images need 36 / 4 = 9 requests. Billing is per image, so the call count changes how long you wait and how many retries you handle, not the price.

Read both ranges from GET /v1/images/models before you hard-code them. The Image API docs say a model accepts only the values its descriptors list, and a request that sets an unlisted parameter returns 400 unsupported_parameter.

Cost by resolution

Sume bills the provider list price times 1.25. For Nano Banana 2.1 the provider tiers in the Sume pricing tables are $0.06 at 0.5K, $0.08 at 1K, $0.12 at 2K and $0.16 at 4K. The table shows the billed price and the 36-image total.

Against GPT Image 2.5

ChatGPT Image 2.5 (openai/gpt-image-2.5) takes up to 16 references, but its price depends on quality and size. A one-reference edit with no quality or size set quotes $0.28175 on Sume (provider $0.2254 x 1.25), because an omitted quality is high and an omitted size reserves the upper bound. 36 of those is 36 x $0.28175 = $10.14.

Setting a size and a lower quality brings that down a lot. See GPT Image 2.5 with no image_size for the effect of the size field.

Keep the shoe the same

Prompt for the one change only: say which part to recolor and that the shape, laces, sole and lighting stay identical. Pass aspect_ratio: "auto" on edits so the output matches the source photo; the docs note that omitting the field is not the same as auto. Reference URLs must be public HTTPS.

Check the output before you ship 36

Do not generate all 36 first. Run one call with n: 4, check that the sole change is the color and that the sole is not drifting, then run the other eight. A bad prompt at $0.40 per call (4 x $0.10) is cheaper to catch on call one than on call nine.

Name the colorways in a list in your own code and put one color in each prompt. Four images from one prompt give you four takes on a single color, which is useful for picking the best one, but it does not give four different colors. For 36 distinct colorways, make 36 single-image calls, or 9 calls that each name four colors and accept that the model decides which image gets which color. Either way, 36 images at 1K bill $3.60.

Keep the source photo at a neutral angle with even light. Edits inherit shadows and reflections from the reference, and a strong colored reflection in the source will tint every variant.