SALD: Self-Referenced Advantage Learning for Diffusion Models
SALD trains a diffusion model against its own easier denoising path, without adding a teacher model or changing inference.
The method compares the same image-caption pair at two noise levels: a lower-noise reference path and a higher-noise training path. It turns the gap between those errors into weights for the harder objective, with extra mechanisms for tracking difficulty over time and weighting frequency-level residuals. The authors report consistent generation-quality gains across several architectures and datasets, with ablations separating the contribution of each component. Source: HF Daily Papers' note.
The method compares the same image-caption pair at two noise levels: a lower-noise reference path and a higher-noise training path. It turns the gap between those errors into weights for the harder objective, with extra mechanisms for tracking difficulty over time and weighting frequency-level residuals. The authors report consistent generation-quality gains across several architectures and datasets, with ablations separating the contribution of each component. Source: HF Daily Papers' note.
score 4