Purify Before You Align: Improving Representation Alignment for Diffusion Models
Author1, Author2, Author3, Author4, Author5
Abstract
This paper introduces local appearance purification (LAP) to enhance representation alignment in diffusion models, improving generation quality by focusing on useful information for generation.
Reality Card
LAP significantly improves generation quality across multiple pretrained representations, achieving a reduction in FID from 53.1 to 42.3 for MAE on ImageNet-256.
LAP improves generation quality with a Spearman correlation of 0.94 predicting the benefit of purification across 13 encoders.
The method's effectiveness may vary with different encoders, which could affect reproducibility across diverse applications.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.