How to edit images with ImageGPT v3

Edit an image with a text instruction using up to five references.

Quality

High quality gives the most detail and takes 30–50 seconds. Low quality (the default) is faster (about 15–20 seconds) and cheaper. In this mode the result always comes back in 2K, even if you select 1K, and the price does not depend on resolution.

References

Upload between one and five images and refer to them in the prompt in upload order as image 1, image 2, ... image 5. Usually image 1 is the main image to edit, and the others supply an object, outfit, background, material, color, or style.

For every reference, state explicitly what to take from it: replace a subject, combine elements, or transfer a style.

Writing a prompt

A strong instruction follows the pattern action → object → desired result. Actions are add, remove, replace, modify, or take from a reference. Objects are a character, item, background, lighting, or color.

Keep it short and unambiguous. Name the exact object instead of using “he,” “she,” “it,” or “that one.”
If everything else must stay the same, say so: “keep the pose and facial expression unchanged.”
Put exact on-image text in double quotes.
Avoid stacking adjectives: a short, precise description works better than a long one.

To point at an exact spot, draw a box, arrow, or circle on the image before uploading it, then name the mark in the prompt and ask to remove it: “Add a TV inside the red box. Remove the marks.”

Examples

Replace the clothing of the person in image 1 with the outfit from image 2. Keep the face, pose, and
background unchanged.

Place the character from image 1 into the background of image 2 in the style of image 3.

Using the material from image 1 and the color from image 2, modify the sofa in image 3.

Replace the largest bread man with a croissant man, keeping the pose and expression unchanged.

Remove the girl's hat. Keep everything else unchanged.

Change the sign text to "OPEN 24/7", preserving its color, typeface, and perspective.

Settings

Choose 1K or 2K resolution and an aspect ratio: auto, 1:1, 3:4, 4:3, 9:16, or 16:9. A single run can produce up to four results.