Everyone (except Zuckerberg and the woman just to the right of him) is consistently looking at something to the right of and slightly above the camera with an uncomfortable/puzzled/mildly dissociating expression. That sort of coherence around something unusual is difficult for genAI, because there likely won’t be many image-label pairs in the training dataset that say “everybody is looking uncomfortable and to the right of the camera”, and because AFAIK there isn’t even an image generating equivalent of the ‘meta-reasoning scratchpad’ that modern LLM implementations have.
Everyone (except Zuckerberg and the woman just to the right of him) is consistently looking at something to the right of and slightly above the camera with an uncomfortable/puzzled/mildly dissociating expression. That sort of coherence around something unusual is difficult for genAI, because there likely won’t be many image-label pairs in the training dataset that say “everybody is looking uncomfortable and to the right of the camera”, and because AFAIK there isn’t even an image generating equivalent of the ‘meta-reasoning scratchpad’ that modern LLM implementations have.