ImageReward and HPSv2 are trained on annotator preferences for generated images; they are used both to rank outputs and to fine-tune models. Do they favour saturated, centred stock-style images over cinematic or graphic-design looks, and are there benchmarks for that shift?
Why this is still open: READMEs report agreement with annotators on their own test sets; out-of-domain behaviour was not documented in the lines read.