GPT Image 1 Edit: change one locked still (4K, 2cr)
GPT Image 1 Edit is an edit. Point at the object. A new text-to-image throws away the still you approved.
GPT Image 1 Edit is an edit. Point at the object. A new text-to-image throws away the still you approved.
GPT Image 1 Edit is OpenAI's instruction-based image row: edit-image and image-to-image, image required, 4K ceiling, 2 credits. The catalog description is the job: describe the change in plain English and it repaints only what you asked, preserving the rest of the frame. Suited to precise product and layout touch-ups. That is not a second generate. If you re-prompt the hero from scratch, you do not have an edit. You have a new still that happens to be on the same brief.
Plain English is the mask
You do not need a painted mask to start. You need a sentence that names the object and the change: swap the label, kill the glare on the lid, move the bottle left of the glass. The AI photo editor is the surface. The instruction is the selection.
A prompt that re-describes the whole scene is how the model "helps" by inventing a different table. Image-to-image work only holds if the instruction is smaller than the frame. Point. Do not restage.
The still has to go in
This row requires an image. Text alone is the wrong family. Lock the still first — photograph, or a text-to-image plate you already signed off — then spend the 2 credits on the delta. Retrying a still is cheap. Throwing away an approved pack because you regenerated "the same shot" is how ecommerce calendars slip.
The record accepts up to six references. That budget is for the still plus supporting views of the same object or layout, not six unrelated concepts. Extra images that disagree with the locked frame are how identity leaks.
4K at 2 credits is the cheap pass
Two credits is the catalog price. Max output resolution is 4K, with SD and HD also listed. Aspect options are 1:1, 2:3, 3:2, 16:9, and 9:16. Styles listed are Realistic, Artistic, and Anime. None of those knobs mean "start over." They mean finish the plate you already have.
Use this when the client approved the composition and you still owe a legal line, a colorway, or a missing SKU. Do not use it as a mood-board engine. Mood boards are a text-to-image row.
One change, then stop
The failure mode is stacking five instructions in one generate so the model averages them. Change the cap color. Look. Then change the reflection. Look. The image-to-image editing workflow is the same discipline at pack scale. This page is only GPT Image 1 Edit: 4K, 2 credits, image in, one object named.
When the still is locked and the pack is approved, motion is a different generate. Do not buy video until this row has nothing left to point at.
FAQ
Can I run GPT Image 1 Edit from text only?
No. The row requires an image. A text-only prompt belongs on a text-to-image model, which will not preserve the still you already approved.
What does "preserving the rest of the frame" actually mean?
The catalog job is instruction-based editing: you name the change, the rest of the layout should hold. If you rewrite the whole scene, you are asking it to generate, not edit.
Why six references?
The record lists a max of six. Use them as supporting views of the same locked object or layout. Do not treat the slots as a collage of unrelated products.
Is 2 credits a 4K-only price?
The catalog price for this generate is 2 credits. Max output resolution is 4K, with SD and HD also listed. 4K is the ceiling, not a reason to throw the still away and start again.