Figure2. TheoverallpipelineoftheproposedFinPercep-RMandCo-evolutionaryCurriculumLearning(CCL)framework. FinPercep-RM
producesaFine-grainedPerceptualDegradationMapthatcapturesspatiallylocalizeddefectlikelihoodandintensity, andtherewardmodel
is progressively expanded from small-variance, coarse-grained rewards to large-variance, fine-grained signals. During training, under the
CCL mechanism, the Generator first learns with the coarse global reward from RM for a stable and easy initialization, and is then co-
0
evolutionarily guided by increasingly strict RM versions to enhance local fidelity and suppress reward hacking.
k