Stress test: GPT-Image 2.5 survives 5 rounds of repeated AI edits

rms80 · x · 2026-09-18

alecwilcock stress-tested the newly released GPT-Image 2.5 Sunburst: iterative editing on the same photo (change the t-shirt, feed output back as input) to see how many rounds it survives.
Result: even at round 5 quality holds up; other models typically degrade into nightmare fuel before round 3.
The model's headline improvement is precise repeated edits without altering the rest of the image.
Commenters ask whether staying in latent space (avoiding round-tripping through real images) could further fix degradation.

More from Multimodal