GPT Image 1.5: change one locked still (4K, 2cr)
GPT Image 1.5 is an edit. Point at the object. A new text-to-image throws away the still you approved.
GPT Image 1.5 is an edit. Point at the object. A new text-to-image throws away the still you approved.
GPT Image 1.5 is OpenAI's row for conversational prompt understanding across text-to-image, image-to-image, and edits. It is strong at following long, specific instructions and rendering legible in-image text. Content type is image. Max output is 4K. Display price is 2 credits. It does not require an image to start a generate — but the job this page is about is the edit: you already have a locked still, and you want one object changed without rolling a new plate.
Point at the object, keep the 4K plate
The approved still is the contract. Color of the bottle, type on the label, the crop the client signed. A fresh text-to-image from the same prompt is a new sample. It will be 4K and it will be close and it will not be the file you approved.
Edit mode is the instruction against that file: swap the mug, fix the typo, drop the extra hand. Conversational prompting is the feature — long, specific instructions, not a three-word restyle. In-image text is the other reason this row exists: if the locked still has lettering, a new generate is how you lose the words you already liked.
The door is the AI photo editor. The image-to-image glossary term is the job name. The ranked edit list is best image editing model.
Text-to-image is the other door on the same row
This model is not edit-only. Categories include text-to-image. You can start from a blank prompt and get a 4K still at 2 credits. That is a valid first plate. It is the wrong button once the plate is locked.
The failure mode is muscle memory: "run it again, but make the box red." Again, from text, is a new still. The red box will sit in a different room. Use the editor path. Point at the box.
Aspects on the row are 1:1, 2:3, and 3:2. Qualities include SD, HD, and 4K. If the locked file is 4K, stay on 4K. Do not "preview" an edit at SD and hope the 4K pass matches. The instruction is cheap. The mismatch is not.
Max reference is 16. That is a stack of stills you can talk about, not a reason to regenerate the hero.
Two credits to change one thing
2 credits is the display price. It is cheap enough that teams treat GPT Image 1.5 as unlimited drafts. Drafts are how you find the still. After approval, spend the 2 credits as an edit, not as a lottery.
The text-to-image tool is for the first plate. The OpenAI roster also has GPT Image 2 and Sora. Those are other rows. This page is 1.5, 4K, 2 credits, conversational edit.
The longer loop — isolate the object, write the instruction, refuse a full regen — is in the image-to-image editing workflow. This page is only the routing rule: if the still is locked, you are in edit.
Do not animate the unapproved plate
Motion is a later row. Image-to-video will treat whatever you upload as the contract. If you "fixed" the label by regenerating the whole still, the video will faithfully animate the wrong bottle. Edit first. Approve again. Then buy motion.
GPT Image 1.5 does not list durations because it does not make clips. Audio is null. There is no soundtrack to write. There is a 4K still to keep.
FAQ
Is GPT Image 1.5 only an editor?
No. The catalog lists text-to-image, image-to-image, and edit-image. Use text-to-image to find the plate. Use edit once that plate is locked. This page is about the second job.
Why not just re-prompt the whole still?
A new text-to-image throws away the still you approved. Layout, type, and product geometry drift. Point at the object. Keep the file.
What resolution does an edit come back at?
The row tops out at 4K, with SD and HD as quality options. Match the locked plate. Do not mix a 4K approval with an SD preview and call them the same still.
Does this replace an image-to-video model?
No. There is no duration and no audio. Lock the 4K still here at 2 credits, then pick an image-to-video row that requires that still.