Learning Discriminative Geometry for Drifting Models
Persistent representation learning improves one-step Drifting Models, cutting pixel-space FID by about 82-95%.
The paper analyzes Drifting Models, which shift iterative distribution refinement from inference to training for one-step generation. It traces poor pixel-space results to representation geometry that controls sample weighting in kernel density estimation, and introduces persistent representation learning that updates that geometry as the generator evolves. The authors also show a current-step gradient equivalence between the KDE ratio loss and drift regression under matched conditions. Across datasets, the method cuts FID by roughly 82-95% versus original pixel-space Drifting Models without pretrained encoders, with further gains from adapted pretrained features and velocity clipping.
- Attributes drifting quality to discriminative geometry used in KDE.
- Learns representations from pixels without pretrained encoders.
- Reduces FID by about 82-95% versus pixel-space drifting.
- Pretrained features and velocity clipping provide further gains.
From the original article
1 min read at huggingface.co
Recently proposed Drifting Models shift iterative distribution refinement from inference to training, enabling effective one-step generation. However, their performance on complex image datasets depends strongly on the representation used to construct the drifting field: pixel-space drifting performs poorly, whereas pretrained feature spaces substantially improve sample quality for reasons that remain unclear.
By Doudou Zhang, Wenwen Hou, Yilin Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.