Attention Sinks in Diffusion Transformers: A Causal Analysis

Brian Summa; Fangzheng Wu

arxiv: 2605.09313 · v3 · pith:CUQ74TIXnew · submitted 2026-05-10 · 💻 cs.CV

Attention Sinks in Diffusion Transformers: A Causal Analysis

Fangzheng Wu , Brian Summa This is my paper

Pith reviewed 2026-05-13 05:55 UTC · model grok-4.3

classification 💻 cs.CV

keywords attention sinksdiffusion transformerstext-to-image generationcausal interventionsemantic alignmentStable Diffusion 3attention mechanisms

0 comments

The pith

Attention sinks in diffusion transformers can be removed without degrading text-image alignment.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines the role of attention sinks in text-to-image diffusion transformers by identifying tokens that receive the most attention at each step and suppressing their influence. Through training-free paired interventions on attention scores and values, the authors test whether these sinks are essential for producing images that match the input text. They show that removing the single strongest sink leaves standard alignment metrics like CLIP-T and most preference scores unchanged across hundreds of prompts. Stronger removals start to impact one preference metric but not others, and the visual changes are much larger than those from random masking at the same budget. This points to a gap between how much the generation process is altered and whether the final image still conveys the intended meaning.

Core claim

Dynamically identifying and suppressing attention sinks per timestep in diffusion transformers via paired interventions on the score and value paths does not degrade text-image alignment measured by CLIP-T or preference proxies at k=1, with only HPS-v2 showing a boundary at k greater than or equal to 10, while the induced perceptual shifts are approximately six times larger than equal-budget random masking, indicating a dissociation between trajectory perturbation and semantic alignment.

What carries the argument

Dynamically identifying dominant attention recipients per timestep and suppressing them via paired training-free interventions on the attention score and value paths.

Load-bearing premise

The interventions isolate the effect of attention sinks without causing other unintended changes in the diffusion trajectory, and the chosen metrics reflect semantic alignment separately from perceptual style.

What would settle it

If suppressing the single top attention sink at each timestep on the 553 GenEval prompts caused a clear drop in CLIP-T scores or consistent degradation in ImageReward, that would contradict the claim that sinks are dispensable for alignment.

Figures

Figures reproduced from arXiv: 2605.09313 by Brian Summa, Fangzheng Wu.

**Figure 1.** Figure 1: Attention sinks in diffusion transformers. (a) In autoregressive LMs, attention sinks often act as stable anchors that attract dominant attention mass. (b) In diffusion transformers, dominant recipients vary across denoising timesteps; we perform a causal test by dynamically identifying sink tokens per step and suppressing them during inference. (c) Sink suppression preserves semantic alignment and prefere… view at source ↗

**Figure 2.** Figure 2: Attention sink dynamics across layers and denoising timesteps. (a) Maximum incoming mass peaks at Layer 12 and decreases during denoising. (b) Attention entropy is inversely correlated with attention concentration (MaxMass). (c) Top-5 concentration increases over time. (d) Maximum activation follows a similar pattern to MaxMass. attention recipients in SD3 are not at index-0. The overlap between dynamic t… view at source ↗

**Figure 3.** Figure 3: Summary of all experimental results. Heatmap showing paired ∆ values across all interventions and metrics. All effects are small and lie within a narrow range around zero; most are statistically non-significant, and all remain within the predefined practical equivalence margin (|∆| < 0.002 for CLIP-T), indicating no meaningful quality degradation from sink removal under any tested condition. B. Extended Re… view at source ↗

**Figure 4.** Figure 4: Dose–response curves for score-path (left) and value-path (right) interventions. Error bars indicate 95% bootstrap CIs. Both curves are flat across the entire intervention range, with all CIs covering zero. D.1. Score-Path Sweep Tables 10 and 11 report the full score-path dose–response under CLIP-T and ImageReward respectively. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

**Figure 5.** Figure 5: reports the index-0 sink ratio before and after full score-path suppression. Baseline ( = 1.0) Intervention ( = 0) 0 1 2 3 4 5 Sink ratio (%) 44,059× (index-0 sink ratio) 4.41% 0% N=32 prompts Proxy verification (index-0 sink ratio) [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

**Figure 6.** Figure 6: Robustness analysis. (a) Multi-layer intervention (L6+12+18). (b) Phase-specific intervention. (c) Cross-architecture validation (SD3 vs SDXL). All conditions show no significant quality degradation. F. Robustness and Attribution Analyses This section provides full experimental details for the robustness and attribution analyses summarized in Section 3.6. F.1. E1: Task-Type Robustness Using 553 GenEval pro… view at source ↗

**Figure 7.** Figure 7: Qualitative drift comparison at k=1 (union-budget protocol, layer 12). Each row shows the same prompt and seed under three conditions: baseline (left), sink-removed (middle), and equal-budget random-removed (right). Per-image annotations report LPIPS and ∆CLIP-T relative to baseline. Sink masking consistently produces larger appearance changes (layout, color, viewpoint) than random masking while preserving… view at source ↗

**Figure 8.** Figure 8: Qualitative drift comparison at k=5. Same layout as [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗

read the original 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.}

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows you can suppress attention sinks in diffusion transformers at low strength without dropping CLIP-T or preference scores, but the perceptual shifts are real and the proxies may not cleanly separate semantics from style.

read the letter

The main takeaway is that attention sinks in these models are not load-bearing for the alignment metrics the field usually tracks. The authors identify dominant tokens per timestep in Stable Diffusion 3, then apply paired training-free interventions on the score and value paths to suppress them. At k=1 the CLIP-T, ImageReward, and HPS-v2 numbers stay flat across 553 GenEval prompts, while the same budget of random masking produces smaller perceptual changes. They also run a quick SDXL check and release the code. That dissociation between trajectory perturbation and metric stability is the concrete new observation, and it is worth having on record for anyone tuning DiT attention patterns. The work is straightforward empirical work with public artifacts, which is the right way to do this kind of targeted intervention study. The soft spot is exactly the one the stress-test note flags: the claim that semantic alignment is preserved rests on the assumption that CLIP-T and HPS-v2 are insensitive to the low-level style shifts the suppression actually produces. The fact that HPS-v2 starts to move at k>=10 while CLIP-T does not already shows the proxies are not interchangeable, so the lack of drop at k=1 does not automatically mean the generated images are semantically equivalent. The sink-identification rule and the exact mechanics of the paired intervention are also described at a level that leaves some room for post-hoc choices. This is the kind of paper that belongs in a reading group focused on generative model internals or efficiency tweaks. It is not a broad theoretical advance, but the experiment is clean enough and the result specific enough that a serious editor should send it out for review rather than desk-reject. The authors have done the basic due diligence of multiple prompts and released code, so referees can check the details directly.

Referee Report

2 major / 2 minor

Summary. The manuscript claims that attention sinks in diffusion transformers for text-to-image models can be causally analyzed by dynamically identifying high-attention tokens and suppressing them through training-free paired interventions on the score and value paths. Using 553 GenEval prompts on Stable Diffusion 3 (with SDXL validation), the authors show that mild suppression (k=1) leaves CLIP-T, ImageReward, and HPS-v2 scores intact, while stronger suppression (k≥10) causes HPS-v2 to degrade in a metric-dependent manner. Perceptual changes from sink suppression are approximately 6 times larger than those from equivalent random masking, suggesting that sinks primarily affect low-level perceptual features rather than semantic alignment.

Significance. This work offers valuable empirical insights into the role of attention sinks in diffusion transformers, extending observations from language models to generative vision models. The large-scale evaluation across hundreds of prompts, corroboration with SDXL, public release of code, and use of multiple alignment and preference metrics provide a solid foundation for the findings. If the central dissociation holds, it implies that attention sinks are not critical for preserving text-image semantic correspondence in DiTs, which could inform future architectural designs and attention mechanism analyses in diffusion models.

major comments (2)

The criteria for dynamically identifying dominant attention recipients per timestep and the exact mechanics of the paired interventions on score and value paths are not described in sufficient detail. This is load-bearing because the causal isolation of sink contributions depends on these choices being free of post-hoc selection or uncontrolled side effects on the diffusion trajectory, as hinted by the limited abstract description and the weakest assumption regarding intervention side effects.
The metric-dependent boundary at k≥10 (HPS-v2 degrades while CLIP-T remains robust) combined with the ~6× larger perceptual shifts for sink-specific suppression versus random masking questions whether the proxies (CLIP-T, ImageReward, HPS-v2) faithfully measure semantic alignment independent of low-level perceptual/style changes. This directly impacts the central claim of preserved alignment under sink removal.

minor comments (2)

The abstract and results would benefit from an explicit table or figure comparing all metrics (CLIP-T, ImageReward, HPS-v2) across k values for both sink suppression and random masking baselines to allow direct visual assessment of the dissociation.
Clarify in the methods how the 553 prompts were sampled from GenEval and whether category balance was enforced, as this affects generalizability of the no-degradation claim at k=1.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and detailed review. The comments highlight important areas for clarification on methodology and metric interpretation. We address each major comment below and will revise the manuscript to strengthen the presentation of our causal analysis.

read point-by-point responses

Referee: The criteria for dynamically identifying dominant attention recipients per timestep and the exact mechanics of the paired interventions on score and value paths are not described in sufficient detail. This is load-bearing because the causal isolation of sink contributions depends on these choices being free of post-hoc selection or uncontrolled side effects on the diffusion trajectory, as hinted by the limited abstract description and the weakest assumption regarding intervention side effects.

Authors: We agree that greater detail is needed for reproducibility and to fully address concerns about side effects. In the revised manuscript, we will expand Section 3.2 with: (i) explicit definition of sink identification as the top-k tokens by mean attention weight across heads at each timestep t (with k selected per prompt based on attention distribution); (ii) precise intervention mechanics—on the score path, we add a large negative bias (-1e6) to the logits of sink tokens prior to softmax; on the value path, we zero-mask the corresponding value vectors. We include pseudocode and a diagram of the modified attention block. To demonstrate absence of uncontrolled side effects, we add controls showing that the noise prediction MSE remains comparable to baseline (within 2%) and that random-token interventions of equal budget produce smaller perceptual shifts. These additions directly support the causal isolation claim. revision: yes
Referee: The metric-dependent boundary at k≥10 (HPS-v2 degrades while CLIP-T remains robust) combined with the ~6× larger perceptual shifts for sink-specific suppression versus random masking questions whether the proxies (CLIP-T, ImageReward, HPS-v2) faithfully measure semantic alignment independent of low-level perceptual/style changes. This directly impacts the central claim of preserved alignment under sink removal.

Authors: We acknowledge the potential for metric confounding and will revise the discussion to explicitly address it. CLIP-T is used as the primary semantic alignment proxy because it is a standard, training-free measure of text-image correspondence in CLIP space; its stability even at k=10 supports that high-level semantics are not disrupted. HPS-v2 and ImageReward degradation at higher k is consistent with their sensitivity to low-level aesthetics and style, which aligns with our finding that sink suppression induces sink-specific perceptual shifts (LPIPS distance ~6× larger than random masking). We will add a new paragraph discussing proxy limitations, report correlations between LPIPS and each metric, and note that the observed dissociation (robust CLIP-T alongside larger perceptual change) is itself evidence that sinks primarily modulate low-level features rather than semantic content. No new experiments are required, but we will clarify this interpretation to avoid overclaiming independence from perceptual effects. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical causal analysis with external benchmarks

full rationale

The paper conducts an empirical causal analysis by dynamically identifying attention sinks in diffusion transformers and applying training-free paired interventions on score and value paths, then measuring impacts on independent external metrics (CLIP-T, ImageReward, HPS-v2) across GenEval prompts with comparisons to random masking. No derivation chain, first-principles result, fitted parameter, or prediction is claimed or present; the central findings are direct experimental observations of metric stability at k=1 and sink-specific perceptual shifts. No self-citations, ansatzes, or uniqueness theorems are invoked as load-bearing elements. The work is self-contained against external benchmarks and does not reduce any claimed result to its inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The analysis rests on standard transformer attention mechanics and the validity of proxy metrics; no new free parameters are introduced beyond the intervention strength k, which is varied explicitly rather than fitted.

free parameters (1)

intervention strength k
Discrete levels (k=1 and k≥10) chosen to test boundary effects; not fitted to optimize any reported outcome.

axioms (1)

standard math Standard scaled dot-product attention operates as defined in the transformer literature
The interventions act on score and value paths of the existing attention mechanism.

pith-pipeline@v0.9.0 · 5481 in / 1244 out tokens · 44228 ms · 2026-05-13T05:55:24.296652+00:00 · methodology

Review history (2 revisions) →

Attention Sinks in Diffusion Transformers: A Causal Analysis

Core claim

What carries the argument

Load-bearing premise

What would settle it

discussion (0)