How to use ImageGPT v3.7

Precise photo editing from a detailed description using up to seven images, up to 16 megapixels.

Images

Upload between one and seven images. The first one is the image you edit. The others are references: a face, an outfit, an object, a background, or a style.

In the request, refer to the images in upload order as @Image1, @Image2 … @Image7.

Prompt enhancement

ImageGPT v3.7 is most precise with a detailed description of the result and does not always get a short request right: it may remove things next to the object or paint the request's own words into a sign. That is why there is an Enhance prompt button under the request field.

  1. Upload the images.
  2. Write a short request: "Remove the cat from the chair and put the lamp from @Image2 on the table."
  3. Click Enhance prompt. The service looks at the uploaded photos and rewrites the request into a detailed description: what each photo shows, what changes, and what stays as it was.
  4. Check the text, adjust it if needed, and start the edit.

Enhancement costs $0.01 and keeps the language of your request. The description contains labels like <Subject 1>: the model uses them to tell objects apart, so leave them in.

Masked editing

By default the model edits the whole image. To change only part of it, switch to By mask and paint the area on @Image1. The model redraws only that area; everything else stays as it was in the original.

  • Paint a little beyond the object so the new part blends in with the light and edges around it.
  • The other images still work as references: "Replace the bell in her hand with the mug from @Image2".
  • The mode sets how detailed the painted area is, not the size of the result.

Writing a request

If you write the request yourself, without enhancement:

  • Describe the finished result, not the process: "she wears a yellow raincoat," not "try changing the coat."
  • Name what must stay unchanged: the face, the pose, the background, the lighting, nearby objects.
  • For each reference, say what to take from it: "the hat from @Image2," "the dress in the color of the vase in @Image3."
  • Put sign or poster text in double quotes, word for word, and add that there is no other lettering.

Examples

Replace the bell in her hand with a red ceramic mug.
Seat the dog from @Image2 on the sofa between them.
Dress her in the hat from @Image2 and the raincoat from @Image3.
Add a neon sign that reads "PROBLEMBO CAFE" on the wall behind the counter.
Turn the photo into a watercolor illustration.

For requests like these, click Enhance prompt first and then start the edit.

NSFW style

The model makes explicit edits without a style, but it often softens scenes with several people: it leaves clothing on or makes poses more restrained. The NSFW style helps with such scenes. Regular edits do not need it, so keep "No style" for them. The style does not change the price or the processing time.

Mode

  • Economy is the default: about 4 megapixels, the fastest and cheapest mode.
  • Standard renders about 8 megapixels.
  • High renders about 16 megapixels. It shows skin, hair, and fine details best and takes a little longer than Standard.

A single run can produce up to two variants.

Aspect ratio

With Auto the result keeps the proportions of the first image. Pick 1:1, 3:4, 4:3, 9:16, or 16:9 for a different frame: the model extends the scene at the edges. When the ratio is far from the photo's, such as a landscape photo in 9:16, the model may recompose the scene and drop minor objects. The price does not depend on the aspect ratio. In By mask mode there is no aspect ratio to choose: the result keeps the size of the original.

Limitations

  • When many people from different photos go into one scene, the model may leave some of them out. Add one or two people per run.
  • Small or long text may come out with mistakes; short quoted text renders well.