Change the print on a t-shirt in a model photo with a design reference

To change the print on a t-shirt in a model photo, send the model photo as the first input_references entry and the new graphic as the second to POST /v1/images with model: "ideogram/ideogram-v4.5", and say that only the print changes. At Sume's medium price that is $0.075 per photo, $0.0375 at low and $0.275 at high. With references, the first image is the one that is edited and up to four more act as references.

The goal is a believable shirt: the print follows the folds, keeps its scale and does not touch the face, hair or hands. That makes the check as important as the call.

Set up the two images

The model photo should show the shirt front on, with the old print visible. The design reference should be the artwork on a plain background, cropped tight and at a decent size. If your design is a wordmark, give the exact text in the prompt too, so the model has two sources for the letters.

Describe what happens to the old print: replace it, same position and approximately the same size, following the fabric folds. Without that, a model may add the new print next to the old one.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ideogram/ideogram-v4.5",
    "prompt": "Image 1 is a model photo. Replace the print on the t-shirt with the artwork from image 2, same position and size, following the fabric folds. Do not change the face, hair, hands, background or the shirt colour.",
    "input_references": [
      {"type": "image_url", "image_url": {"url": "https://example.com/model-shirt.jpg"}},
      {"type": "image_url", "image_url": {"url": "https://example.com/new-print.png"}}
    ],
    "quality": "medium"
  }'

Edit or mockup

There are two ways to make this picture. A flat mockup places the design on a blank shirt template with your own code, and it gives exact artwork but a flat look. An AI edit gives a natural fall of fabric but may alter the artwork. Pick based on how closely the print has to match the file: an exact brand mark is a mockup job, a fashion lookbook is an edit job.

For an edit where the print has to be exact, use a mask on GPT Image 2.5 to confine the change to the chest area. It takes up to 16 references and an optional mask_url, and costs $0.0835 at high for a one-reference 1024x1024 edit in Sume's pricing fixtures.

Cost

Ideogram amounts are Sume's per-image prices. The GPT row is the billed amount for a 1024x1024 edit with one input image; a second reference adds some input tokens.

What to verify

Compare the new print with the file at full size. Check spelling in any lettering, the colours, and the relationship to the folds. Then diff the whole photo against the source: only the shirt chest should change. If the model changed a face or a hand, discard the result; those are the details a model should never touch in a lookbook.

A completed edit is billed, and a failed job is not. In a pilot of five photos, count how many pass on the first try, and set your batch budget from that rate instead of from the price of one image.

Common failure modes

The most common failures are a print that is too small, a print that ignores the fold lines so it looks stickered on, and artwork whose colours drift toward the shirt colour. Fix the first with an explicit size, such as covering about half the chest width. Fix the second by naming the folds in the prompt. Fix the third by saying the artwork colours must stay as they are in image 2.

Another failure is a changed shirt: the collar, the sleeve length or the colour of the fabric moves. Say what must stay, and compare the shirt outline against the source. If the outline moves, rerun rather than retouching, since a rerun at low quality costs $0.0375 and keeps the edit honest.

Run a three-photo pilot before a batch. If two of three pass your review, a batch of 30 will need about 45 attempts, and you can budget that at the per-image prices in the table.