Attention Sinks in Diffusion Transformers: A Causal Analysis

TL;DR

This paper investigates the role of attention sinks in diffusion transformers, showing that removing them does not harm core alignment metrics and revealing a dissociation between attention perturbations and semantic quality.

Contribution

It introduces a causal analysis method to identify and suppress attention sinks in diffusion models, demonstrating their limited impact on semantic alignment.

Findings

Removing attention sinks does not degrade text-image alignment at k=1.

Stronger interventions reveal metric-dependent effects on HPS-v2.

Sink suppression causes larger perceptual shifts than random masking.

Abstract

Attention sinks -- tokens that receive disproportionate attention mass -- are assumed to be functionally important in autoregressive language models, but their role in diffusion transformers remains unclear. We present a causal analysis in text-to-image diffusion, dynamically identifying dominant attention recipients per timestep and suppressing them via paired, training-free interventions on the score and value paths. Across 553 GenEval prompts on Stable Diffusion~3 (with SDXL corroboration), removing these sinks does not degrade text-image alignment (CLIP-T) or preference proxies (ImageReward, HPS-v2) at $k{=}1$; only under stronger interventions ($k\!\geq\!10$) does HPS-v2 exhibit a metric-dependent boundary, while CLIP-T remains robust throughout. The perceptual shifts induced by suppression are nonetheless \emph{sink-specific} -- $\sim\!6\times$ larger than equal-budget random masking -- revealing an empirical dissociation between trajectory-level perturbation and \emph{semantic alignment} in diffusion transformers. \footnote{Code available at https://github.com/wfz666/ICML26-attention-sink.}