Learning Discriminative Geometry for Drifting Models
The paper argues that Drifting Models fail in pixel space because the representation geometry drives the KDE weighting behind drift.
The authors propose 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, using that link to justify more direct control of drifting velocity. On tested image datasets, the method learns from pixels without pretrained encoders and cuts FID by about 82-95% versus the original pixel-space Drifting Models. Further gains come from adapting pretrained representations and velocity clipping. ArXiv · AI/CL/LG's note
The authors propose 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, using that link to justify more direct control of drifting velocity. On tested image datasets, the method learns from pixels without pretrained encoders and cuts FID by about 82-95% versus the original pixel-space Drifting Models. Further gains come from adapting pretrained representations and velocity clipping. ArXiv · AI/CL/LG's note
score 4