Tsukada Lab. Computer Network Group

How Do Diffusion Classifiers Decide? A Bias-Centric Evaluation

Saba Fathi, Fardin Ayar, Maryam Abdolali, Ehsan Javanmardi, Manabu Tsukada, Mahdi Javanmardi

, Jul 4, 2026

Summary in three lines

  • ASOB-Bench measures attribute binding, object size and order, and background dependence in diffusion classifiers.
  • The study visualizes reconstruction error and cross-attention to see which regions influence a classification.
  • Tests using natural and unusual object colors expose attribute errors linked to familiar color–object associations.

Figures from the paper

Figure 1: Three ASOB-Bench bias categories
Figure 1. Three ASOB-Bench bias categoriesSource: Figure 1 from Fathi et al., arXiv (2026) (authors’ figure from the manuscript)The examples show how reconstruction error and attention reveal attribute, size and background effects.

How to cite

S. Fathi, F. Ayar, M. Abdolali, E. Javanmardi, M. Tsukada, and M. Javanmardi, “How Do Diffusion Classifiers Decide? A Bias-Centric Evaluation,” in , 2026.

BibTeX

@unpublished{Fathi2026,
  title         = {How Do Diffusion Classifiers Decide? A Bias-Centric Evaluation},
  author        = {Saba Fathi and Fardin Ayar and Maryam Abdolali and Ehsan Javanmardi and Manabu Tsukada and Mahdi Javanmardi},
  year          = {2026},
  date          = {2026-07-04},
  eprint        = {2607.03831},
  archiveprefix = {arXiv},
  primaryclass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.03831},
  note          = {Preprint}
}