GPT Image 2 & 2.5 Image Generator

All three take text, one image to edit or several references, at 1K, 2K or 4K. 2.5 Flare is the default. 2.5 Sunburst costs the same as 2.5 Flare at 1K and 2K and more at 4K (see the table above). Version 2 is the oldest and the cheapest. The look differs between versions, so run the same prompt on two of them and keep the one you prefer. The newest version, 2.6, has its own page: GPT Image 2.6.

Yes. Upload one image as the reference and describe what to keep and what to change, keep-list first: keep her face, hair and jacket; change the background to a night street. Several references can be combined into one picture as well. Only upload photos of yourself or of people who have agreed.

A fixed number of credits per image for each version and resolution; 4K costs the most. See the table above; the Generate button shows the exact price for what you picked. Failed generations are refunded.

Yes. The Free plan includes monthly credits you can spend on any version, and the price of each run is shown before you confirm. Drafting at 1K stretches those credits furthest. Failed or refused runs are refunded automatically; plans and credit packs are on the pricing page.

Most of them. 2.5 Flare powers the image generator AI by default, the mascot generator, the AI outfit changer, AI product photography and UGC product photos. Version 2 powers the character turnaround and the consistent character sheet, where views must match. The character reference sheet runs on GPT Image 2.6 Sunburst.

Yes. 2.5 Flare draws titles, short bios and labels in one image. Our character reference sheets are made in a single 4K image by the newer GPT Image 2.6 Sunburst. Small text needs 4K to stay readable, and you should check the spelling on the result.

Up to 4 per run in the generator on this page. With more than one, name each image's role in the prompt; for a recurring character, one multi-angle sheet does more than several loose photos.

Yes, when the reference shows the face clearly. In our outfit tests every model we tried kept the face, and on character sheets version 2 kept six views of one person consistent. A sheet with several angles holds up better than a single photo when the head turns.

The model has its own content policy and refuses some requests. In our tests it refused a bikini on a photoreal character but accepted a one-piece swimsuit, and accepted the same bikini on a 3D character. A refused run costs nothing; the credits come back automatically.

Yes. Add it as the first frame in image to video and describe the motion, or make a start and an end image for first and last frame video. Generate the still in the video's aspect ratio, since the clip starts from that exact frame.