Add a hand holding the product to a packshot with GPT Image 2.5

To add a hand holding a product to a packshot, send the packshot and a reference of the pose you want to POST /v1/images with model: "openai/gpt-image-2.5", and ask for a natural hand holding the same product with the label facing the camera. ChatGPT Image 2.5 takes up to 16 references on Sume, so you can add a pose image, a skin tone reference and a close-up of the label. A one-reference 1024x1024 edit at high is $0.0835 in Sume's pricing fixtures.

The risks are scale, grip and label. Hands that are too large or too small make a product look wrong, fingers that do not touch the product look pasted on, and labels that rotate out of view lose the brand.

Choose the references

The packshot is the product's ground truth, so use a clean front view at a decent size. A second reference of the product from the side or a label close-up helps if the label is the important part. A third reference for the hand is optional: it can be a photo of a similar grip, but be careful about using a real person's likeness without permission.

Number the references in the prompt and give each a role. Image 1 is the product, image 2 is the grip pose. This is the same pattern as any multi-reference edit, and the docs do not define an order rule for ChatGPT Image 2.5, so spell it out.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Image 1 is the product. Image 2 is a grip pose reference. Show an adult hand holding the exact product from image 1 at chest height, label facing the camera, natural scale, soft window light, plain background.",
    "input_references": [
      {"type": "image_url", "image_url": {"url": "https://example.com/packshot.png"}},
      {"type": "image_url", "image_url": {"url": "https://example.com/grip-pose.jpg"}}
    ],
    "quality": "high",
    "aspect_ratio": "4:5"
  }'

Scale and grip

Give the product a size. A 30 ml bottle is held between thumb and fingers; a 500 ml bottle is held in a palm. Say what the hand does: pinches the cap, cradles the base, wraps around the middle. A vague prompt gets a vague hold, and the product floats near the fingers.

Say what the label shows and which side faces the camera. If the label text matters, include a label close-up as a reference and check the text letter by letter in the result.

Cost by quality

The fixture amounts are for one input image at 1024x1024. A second or third reference adds input tokens, so treat them as a floor.

Workflow that keeps cost down

Explore at low and medium. Ten medium takes are $0.21. Pick the best composition and rerun once at high. Because Sume has no seed for these models, the rerun at high is a new image, so judge the final one on its own merits. Keep every URL, since you may prefer an earlier take.

Check the hand against reality: five fingers, correct thumb, no extra joints, nails pointing the right way. Hands are where models fail first, and a mask-based fix is possible on this model if one finger is wrong; see the related post on masked hand fixes.

Also say what the hand is not: no jewellery, no tattoos, no text on the skin, and no second hand unless you want one. Models like to add details to hands, and every extra detail is another thing to review. Keep the background plain so the product and hand are the only objects to check.

Alternatives

If you need the product pixels to stay exact, compose the photo yourself: cut the product out with RMBG at $0.0225, place it on a stock photo of a hand in your own editor, and add a shadow. That route keeps the label perfect but needs more work and a hand photo you are licensed to use.

If you need variety rather than exactness, the edit route wins. Generate six takes at medium for $0.126 total and choose one. In both cases, never publish a hand-held image that implies a size or a use the product does not have.