some question about the final reward score

[wh0x]

i have some question, when use single_inference, the rewards is a tensor like tensor([[ 1.4358, -1.9688]], device='cuda:0'), and in the inference code, get the reward[0] as final result, what does the tensor value mean ? and why choose reward[0], The paper mentions that the predictions are for the mean and variance. does 1.4358 is mean and -1.9688 is var? and why choose mean as final result, by the way, i am trying this model as a scorer for image edit data quality eval

[KemingWu]

Thanks for the question. For the released EditReward checkpoints, the output tensor has shape [B, 2], where the first value is the reward mean (the main scalar score) and the second value is the uncertainty term in log-space rather than the raw variance itself. So in your example, 1.4358 is the score we use as the final result, while -1.9688 is the log-uncertainty term. That is why the inference examples take reward[0]: it is the primary quality / preference signal for ranking or filtering edited images, while the second value can be viewed as an auxiliary confidence indicator.