GPT-IMAGE-2 self-critiques output, takes 11 minutes

This looks like test-time compute applied to image generation: better prompt fidelity and layout correctness in exchange for a much worse turnaround time. That is useful for final assets, but it is a serious tax on any workflow that depends on rapid iteration.

• –Self-review loops can help with text rendering, object consistency, and prompt adherence on complex compositions
• –Eleven-minute generations are fine for hero images, but too slow for UI mockups, ad variants, and storyboard-style exploration
• –The real product question is whether users get a speed-quality control, or whether the model silently burns latency to improve results
• –Because this is a Reddit report, treat it as anecdotal until OpenAI documents the behavior or others reproduce it reliably