Change one of three identical bottles: position words or a mask

When three bottles look the same, name the one you want by where it stands, and add a mask if the edit must stay inside it. Flux 3 Image, launched on October 1, 2026, solves this with element ids: the Decoder and TechTimes both describe a JSON table in which each element has an id and a box, so "bottle_2" is a specific object. Sume does not list Flux 3 Image, so on Sume the options are the prompt and, for one model, mask_url.

Route 1: position words on Ideogram 4.5

Send the photo as the first entry of input_references. Name the target by its place and by something it does not share with the others: "the middle bottle", "the bottle on the right, nearest the camera". Then say what must not change: "Change the label on the middle bottle to green. Keep the left and right bottles, the table and the lighting unchanged." Count objects from the viewer's side, since the model reads left and right as it sees the picture.

This costs $0.0375 per low-quality image on Sume (the provider list of $0.03 times 1.25), so ten tries are $0.375. It is the cheaper route and often enough when the bottles are spaced apart.

Route 2: a mask on GPT Image 2.5

The Image API docs document mask_url for openai/gpt-image-2.5 only. Build a mask the size of the photo, mark the middle bottle's area, and host it at a public HTTPS URL like the photo itself. The docs do not say which colour means edit, so render one cheap test with the mask and see which bottle changed before you run a full pass.

GPT Image 2.5 bills by tokens: the docs give $30 per million output tokens, $8 per million image input tokens and $5 per million text input tokens, before Sume's margin. At 1024x1024, xhigh output is $0.09366 and max is $0.21072, and input tokens add to that. Pick quality with the table in mind rather than leaving the default of high.

Which to try first

Start with words. If the wrong bottle changes twice in a row, move to the mask.

Sume bills a failed generation at nothing, according to the docs, but a successful image of the wrong bottle is billed like any other. That favours a short test loop: three low-quality tries on Ideogram 4.5 cost $0.1125 in total, less than one max-quality GPT Image 2.5 render at 1024x1024 before input tokens.

What to check

After the edit, crop each bottle from the original and the result and compare the two you did not touch. Identical products make drift easy to miss, because a slightly different label on bottle one looks like the product. A difference image from Pillow (ImageChops.difference) will show it at once. The mask page, GPT Image 2.5 mask_url edits, covers building the mask file.