Cross-scale Alignment Problem in GAN Training Identified

A recent study questions the understanding of multi-stage GAN synthesis as a hierarchical process from coarse to fine. The researchers contend that traditional adversarial supervision by scale does not create an effective hierarchy, as each intermediate image is independently aligned with the real distribution at its specific resolution, failing to guarantee that outputs from different stages correspond to the same sample. Furthermore, the image generated at each stage is not utilized as a clear refinement goal for the next stage, which permits later stages to deviate. This issue is referred to as cross-scale alignment. The findings can be found on arXiv.