How to use ImageGPT v3.5

Edit an image using up to eight references.

References

Upload between one and seven images and refer to them in upload order as <image1>, <image2>, ... <image7>. The first image is the main image and determines the output proportions. The remaining images can define an identity, outfit, pose, background, object, palette, or style.

Writing a prompt

A strong request describes: what to change → where to change it → which reference supplies the details → what must stay unchanged → how lighting and style should match.

Assign an explicit role to every reference.
Name the exact subject and image number instead of using “he,” “she,” or “it.”
List the features that must not change, such as the face, pose, composition, clothing, or background.
Put exact poster or sign text in double quotes and specify its color, typeface, and perspective.
Split complex edits into stages: handle the character and outfit first, then the background and lighting.

Examples

Preserve the face, pose, and body proportions of the person in <image1>. Replace only the outfit
with the outfit from <image2>, accurately transferring its material and details while matching the
lighting in <image1>.

Preserve the subject from <image1>. Replace the background with the scene from <image2> and apply
its lighting naturally to the face, hair, and clothing. Keep everything else unchanged.

In <image1>, replace the sign text with "OPEN 24/7". Preserve the original color, letter weight,
angle, perspective, and glow. Keep the surrounding architecture unchanged.

Resolution

Choose the output resolution: 1K is about 1.6 megapixels (1280×1280 for a square), 2K is about 4 megapixels (2048×2048 for a square). The result keeps the proportions of the first image, so there is no aspect ratio to choose. References larger than 4 megapixels are automatically reduced while preserving their proportions. A single run can produce up to four results.