Possible regression in reference-image-based image edit/regenerate behavior since around Mar 12, 4:00 PM JST
OpenAI Developer Community
March 13, 2026
Additional observation:
I also observed cases where I cropped and enlarged only the area that needed correction, then attached that partial image and requested a localized fix.
Even in those cases, the output did not preserve the input image ratio. A 2:3 cropped input could come back as a 1:1 square output instead.
This suggests the issue is not limited to full-frame reference edits. It may also affect partial or localized edit flows, where non-target areas are effectively discarded and the result is re-emitted on a fixed square canvas.
At least from what I can observe externally, the current behavior does not appear to preserve source dimensions reliably. So far, I have only observed outputs collapsing into fixed formats such as 1024×1536, 1536×1024, or 1024×1024, rather than preserving the original input frame.
I am not claiming this as a confirmed internal mechanism. This is simply the repeated pattern I have observed from the outside.
Discussion in the ATmosphere