Learning Discriminative Geometry for Drifting Models
The paper says drifting models fail in pixel space because the representation geometry misweights samples during KDE-based drift.
The authors introduce persistent representation learning, which updates the representation as the generator changes across batches. They connect the KDE ratio loss to drift regression under matched conditions, arguing for direct control of drifting velocity. On multiple datasets, their pixel-trained approach cuts FID by about 82-95% versus the original pixel-space drifting models, without using pretrained encoders. Further gains come from adapting pretrained representations and applying velocity clipping. HF Daily Papers' note
The authors introduce persistent representation learning, which updates the representation as the generator changes across batches. They connect the KDE ratio loss to drift regression under matched conditions, arguing for direct control of drifting velocity. On multiple datasets, their pixel-trained approach cuts FID by about 82-95% versus the original pixel-space drifting models, without using pretrained encoders. Further gains come from adapting pretrained representations and applying velocity clipping. HF Daily Papers' note
score 4