Possible regression in reference-image-based image edit/regenerate behavior since around Mar 12, 4:00 PM JST

ChatGPT Bugs
,
1

Hello,

I would like to report a possible regression in reference-image-based image editing / regeneration behavior.

Based on repeated tests on my side, I observed a clear change starting around March 12, approximately 4:00 PM JST.

Since that time, when I provide a reference image and request either:

- an image edit / correction, or

- a localized regeneration / partial fix,

the output no longer preserves the original image boundaries and consistently introduces issues such as:

- cropping / trimming,

- framing changes,

- canvas size changes,

- modification of non-target areas.

This happens even when the request explicitly and strictly prohibits any canvas size change, cropping, trimming, reframing, or composition change, and asks for a minimal local correction only.

Observed results from recent testing:

- more than 20 attempts

- cropping / trimming occurred in all attempts

- 0 attempts preserved the original frame

- 0 successful correction results

The issue is still ongoing at the time of writing.

The main concern is not only output quality, but failure of basic reference-image preservation behavior:

- the original canvas is not preserved reliably

- framing changes occur without being requested

- non-target areas are modified during edit / regenerate requests

From my testing, this appears to be a regression in the reliability of reference-image-based editing.

Please investigate whether there was any backend change, model update, routing change, or editing pipeline change around that period.

If necessary, I can provide additional examples of the failure pattern.

Thank you.

2

Additional observation from separate testing:

This additional test was performed using text-to-image generation, not image-to-image editing.

In these separate tests, I observed that portrait-oriented outputs consistently resolved to 1024×1536, while landscape-oriented outputs consistently resolved to 1536×1024.

At this stage, I cannot state this as a confirmed system-wide rule, because testing is still limited. However, across 4 additional tests of this type, together with the earlier 20+ edit / regenerate tests, all observed outcomes so far appear to collapse into one of these two final canvas sizes.

Because of that, one possible interpretation is that the current output path may be forcing results into a fixed portrait or landscape canvas format, rather than preserving the original source dimensions during reference-image-based edit / regenerate requests.

I cannot verify the internal cause from the outside, but this behavior may be related to the current failures in preserving the original frame and canvas.