Persistent representation learning cuts Drifting Models FID by 82–95% from pixels

Doudou Zhang · hf · 2026-10-07

Context: Drifting Models shift iterative distribution refinement from inference to training for one-step generation, but performance hinges on the representation used to build the drifting field — pixel-space fails while pretrained feature spaces excel, for unclear reasons.
Finding: the gap traces to the discriminative geometry of the representation, which governs sample weighting in KDE and thus drift; the authors prove a current-step gradient equivalence between the KDE ratio loss and drift regression loss, motivating direct control of drifting velocity.
Method: persistent representation learning continuously learns a more discriminative representation as the generator evolves across batches.
Results: across datasets, learning from pixels alone reduces FID by 82–95% over original pixel-space Drifting Models, without pretrained encoders; adapting pretrained representations and velocity clipping add further gains.

More from Multimodal