cond-mat.dis-nn, cs.LG, stat.ML

The Interplay of Data Structure and Imbalance in the Learning Dynamics of Diffusion Models

arXiv:2605.06367v1 Announce Type: new
Abstract: Real-world datasets are inherently heterogeneous, yet how per-class structural differences and sampling imbalance shape the training dynamics of diffusion models-and potentially exacerbate disparities-re…