SOI by directly leveraging human feedback without being restricted by a predefined reward model nor cumbersome model selection.

Our FreGAN improves the overall quality of generated images and raises the model's frequency awareness, encouraging the generator to produce precise

CIFAR-10 is an image classification dataset. It contains 50,000 training images and 10,000 test images of 10 classes. We adopt the standard data augmentation strategy on CIFAR datasets, i.e., padding 4 pixels on each side of an image and randomly flipping it horizontally, and Tiny ImageNet is a subset of ImageNet. We adopt the standard data argumentation, i.e., padding 8 pixels on each side ImageNet is a large-scale image classification dataset containing 1.28 million training images and The standard augmentation [6, 4] is adopted. To save the cost, we do a very basic search instead of grid search.

We thank all reviewers for their valuable feedback and constructive suggestions. Major comments are addressed below. Several works (eg, [7] and [11]) follow a similar rationale. We thank the reviewer for suggesting these large-scale image datasets. Q1: What "evidence-based entropy" is when claiming entropy can be decomposed into vacuity and dissonance.