Which image editing models drift over 30 consecutive edits

Artificial Analysis, which runs an image-editing leaderboard that scores single edits, tested what happens when edits pile up. In a post on X on October 9, 2026, it describes giving four models the same photo of a room and a 30-step real estate staging sequence, such as lighting the fire, adding a sofa, repainting the walls and switching day to twilight, with each model editing its own previous output. Ideogram 4.5 kept most of the room intact. The others drifted, most visibly GPT Image 2.5 Sunburst, which is first on the single-edit leaderboard.

The explanation it gives is about how much of the picture each model redraws. Ideogram 4.5 and Black Forest Labs' FLUX 3 edit locally: on small edits such as adding a vase of tulips, Artificial Analysis says it measured that they left 95% or more of the image essentially untouched. GPT Image 2.5 Sunburst re-renders most of the image on every edit, leaving only about a fifth unchanged, so small shifts in colour and detail compound over the turns. Nano Banana 2.1 sits in between: its edits stay local, but the rest of the image shifts slightly and gradually darkens.

That is why a single-edit score can miss the problem: the model that tops it redraws the most each time. Ideogram 4.5 had debuted at #23 on Artificial Analysis's editing leaderboard a week earlier, with Ideogram positioning it to reduce pixel shifts and colour changes across multi-turn edits. This is one evaluator's test of one staging sequence, and the attached video showing the results could not be reviewed, so the account here rests on the post's text.