Learning Discriminative Geometry for Drifting Models

Learning Discriminative Geometry for Drifting Models. By Doudou Zhang and three others, from the ELLIS Institute Finland and Aalto University. Published on the third of October. Its code repository is public but slight: seven source files.

Drifting Models, proposed recently, move the iterative refinement of a distribution out of inference and into training, which lets them generate in a single step. Their weak point is the representation used to build the drifting field. On complex image datasets, drifting in pixel space performs poorly. Drifting in a pretrained feature space substantially improves sample quality, and why that should be so has remained unclear.

The authors trace the gap to the discriminative geometry of the representation. That geometry decides how samples are weighted in kernel density estimation, and the weighting in turn decides the drift. If geometry is what matters, then a good representation need not be borrowed from a pretrained encoder. It can be learned. Persistent representation learning does exactly that, continuously learning a more discriminative geometry as the generator evolves from batch to batch.

There is a theoretical link underneath. Under matched conditions, the authors establish a current-step gradient equivalence between a KDE ratio loss and the drift regression loss. That connects generator optimization based on density ratios to empirical drifting, and it motivates controlling the drifting velocity directly rather than leaving it implicit.

Put the two pieces side by side. The representation is not fixed in advance but trained persistently, keeping pace with a generator whose samples change from one batch to the next. And the equivalence between the two losses means the density-ratio view and the drifting view describe the same update at the current step, which is what licenses acting on the velocity itself.

Across multiple datasets, the method learns effective discriminative representations straight from pixels, with no pretrained encoder, and reduces FID by roughly eighty-two to ninety-five percent compared with the original pixel-space Drifting Models. Two further steps add more: adapting pretrained representations, and applying velocity clipping.

A puzzle about why one feature space works and another fails becomes a recipe. Once the advantage is located in discriminative geometry, the geometry can be trained alongside the generator, and pixels stop being the handicap they were.