my conspiracy theory is that OpenAI intentionally adds acutance haloing to images v2 so anyone trying to distill from its outputs overfits to it and then OpenAI drops a better model without it
probably not true but its still what happened with grok imagine 2
Grok Imagine Image 2.0 prompt adherence is way too strict / broken. Even when you prompt extremely hard and specific, it still ignores half the instructions. You can write detailed descriptions for pose, lighting, skin texture, exact details… and it just does whatever it wants
You can’t tell me going from this — Grok Imagine 1.5 (left) — to this — Grok Imagine 2— is an upgrade. Image 2 looks significantly worse (same prompts). fix this shit @elonmusk
why did they get rid of that verison of grok imagine maybe it was a few months ago i think it improved your prompt but the generations that came with it were outstanding. i understand that some people don't want their prompt changed they should make it an optional setting.