Attention of a Kiss: Exploring Attention Maps in Video Diffusion for XAIxArts

Adam Cole; Mick Grierson

REVIEW 3 major objections 4 minor 16 references

Attention of a Kiss: Exploring Attention Maps in Video Diffusion for XAIxArts

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that cross-attention maps in Wan's text-to-video model align with the objects their prompt tokens name, frame by frame, and can serve artists as both an interpretative tool and a raw material.

desk verdict A modest but honest arts-first application of cross-attention visualization to Wan2.1; the interpretability claim outruns the evidence, but the tool and case study are worth a look. read the letter →

arxiv 2509.05323 v2 pith:I6JODFUP submitted 2025-08-30 cs.AI cs.MM

Adam Cole , Mick Grierson This is my paper

classification cs.AIcs.MM

keywords cross-attentionmapsvideodiffusiontext-to-videogenerationattentionvisualizationexplainableAIWanmodelXAIxArtsart

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the internal attention of a modern text-to-video diffusion model is visible and interpretable, and that artists can use that visibility as creative material. Building on the open-source Wan model, the authors extract the cross-attention map for every prompt token—the softmax-scaled match between the token and regions of the generated video—and render it as a heatmap over time. Their probes show the heatmap for "cat" sitting on the cat, separate tokens separating in multi-object scenes, and a more diffuse but lip-clustering map for the abstract token "kiss." If the claim holds, prompt-driven video generation stops being a black box: artists can watch which words shape which pixels and intervene in the model's internal logic rather than only its outputs. The paper also reports that maps can be noisy or hard to read, so it frames attention maps as an interpretive lens rather than a full explanation.

What carries the argument

The carrying mechanism is the cross-attention map, computed as softmax(QK^T), where queries come from the prompt tokens and keys from the embedded video representation. Each map therefore records how much every token influences every spatial-temporal region of the video. The paper's technical contribution is the extraction wrapper that captures these maps across diffusion steps, blocks, and heads, and the visualisation pipeline that reshapes the flat embedding into a [frames × height × width] tensor and upscales it to match the output video, producing heatmaps that can be inspected per head, block, step, or averaged.

What would settle it

Compare attention-map peaks for a named object, say 'cat', with the object's location in the generated frames using an object detector or segmentation model across many videos. If the heatmap's brightest region routinely misses the cat, drifts off it between frames, or lights up when the cat is absent, the claim that attention maps coherently align with object regions over time is refuted.

Watch

Extended reading notes

Core claim

The central discovery is a method plus an observation. The method is a wrapper around Wan's cross-attention layers that stores the attention maps produced at every diffusion step, transformer block, and attention head; each stored map is then reshaped from the flat latent embedding into a latent video tensor and upscaled to output resolution for viewing. The observation is that the resulting maps are spatially meaningful: for concrete objects named in the prompt, brightness concentrates on the object's actual location in the generated frames, the concentration sharpens over diffusion steps, and different tokens separately occupy different scene regions. For abstract tokens, the correspondenc

Load-bearing premise

The core assumption is that flattening and upscaling the attention maps preserves where each prompt token actually acts in the output frames—so that a bright heatmap spot really sits on the object the token names—and this spatial faithfulness is never tested quantitatively.

Editorial extensions

If this is right

Artists can inspect exactly which prompt tokens control which video regions, at fine detail per attention head or broad overview averaged across steps and blocks.
Prompt-steered video generation becomes more legible: users can see the model establish composition early, before it is visible in the noisy output.
Attention maps can serve as raw artistic material, since the maps themselves carry the temporal structure of the generation.
The technique extends the early-video-art practice of intervening in a medium's internal signal logic to generative AI, pointing toward "network bending."

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same visualisation pipeline could be used as a diagnostic for prompt-object binding failures: if the attention peak and the object's segmentation in the output disagree, that flags where the generative model lost the referent.
Because attention appears to lock onto object regions early in the diffusion process, the extracted maps could be combined with latent-space editing to steer or correct composition at the point where the decision is actually made.
A quantitative test is within reach: compute attention-map centroids for a token and compare them across frames with object tracks from an off-the-shelf detector, turning the paper's qualitative "coherent alignment" into a measurable claim.
The paper's own "network bending" suggestion implies a broader art practice where attention maps are not just visualised but modified during generation, effectively making prompt-to-region binding a creative parameter.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a short, honest workshop paper that ports known cross-attention visualization to a video diffusion transformer (Wan2.1) and packages it as an artist's tool. The technical novelty is thin—they cite prior work doing attention analysis in video diffusion [8,16]—and the central claim that attention maps 'coherently align with object regions' rests on a handful of selected stills. But for an arts-focused venue, that might be enough, especially since the limitations section is candid about noisy and unintelligible maps.

What's actually new: the specific tool for Wan2.1, the fine-grained views across heads, blocks, and diffusion steps, and the 'Attention of a Kiss' video study that uses the 'kiss' token's attention evolution as raw artistic material. The paper is well written and doesn't oversell—Section 5.2 admits the maps are often noisy, and the conclusion frames attention maps as 'interpretive tools rather than comprehensive explanations.' That's the right tone.

The soft spots are real. The load-bearing assumption is that the flat latent embedding reshapes into [frames, H, W] and upscales to pixel space in a way that preserves spatial correspondence. They never validate that. The evidence for alignment is a few selected examples (cat, soccer ball, Eiffel Tower) assessed by eye, with no quantitative overlap measure, no seed variation, no code release. The paper itself concedes the maps are sometimes unintelligible. So the sentence 'confirming the interpretability of Wan's attention maps' overstates what the data show. A more accurate claim would be 'in several examples, attention maps appear to track the referenced object.'

That said, don't over-correct. For an XAIxArts workshop paper, the standard is different from a top-tier ML venue. The work is a proof-of-concept, and the authors say so implicitly. The missing code and quantitative validation are real limitations, but they don't sink the artistic contribution. If the tool were released and the alignment claim tested with even one quantitative measure (e.g., IoU with segmentation masks across a few seeds), the paper would be much stronger.

Who this is for: people working at the intersection of explainability and arts practice, or anyone wanting to visualize cross-attention in Wan2.1. It's not a technical reference for interpretability methods. I'd send it to peer review—it's a legitimate workshop contribution, and the honest limitations make it a fair candidate for revision rather than rejection.

Referee Report

3 major / 4 minor

Summary. The paper presents a method for extracting and visualizing cross-attention maps from the Wan2.1 text-to-video diffusion transformer, targeting artists and the XAIxArts community. The approach intercepts cross-attention maps during generation, reshapes them from a flat latent embedding to a 3D video tensor, upscales them, and displays them as heatmaps per token, head, block, and diffusion step. The paper reports exploratory qualitative probes—single object, multiple objects, and an abstract 'kiss' concept—and claims that attention maps 'coherently align with object regions over time,' confirming the interpretability of Wan's attention maps. It also presents an artistic video study made from the 'kiss' token's attention maps, and discusses limitations, particularly noisy or unintelligible maps for abstract tokens. The appendix provides detailed settings for the multi-object probe and visualizations across blocks, heads, and steps.

Significance. If the alignment claim holds, the contribution is valuable for artists and XAIxArts: it offers a lightweight interpretability window into a modern open-source video diffusion model, with visual outputs that can serve both analysis and artistic raw material. The paper includes detailed experimental settings for the main multi-object probe, a public video artifact, and an explicit connection to prior network-bending work. However, the central evidence is qualitative and selective: no quantitative evaluation, no seed variation, no validation of the latent-to-pixel correspondence, and no code release. The core technical demonstration—that attention maps can be extracted and viewed—is plausible and reproducible in principle, but the stronger interpretability claim is not yet supported.

major comments (3)

[Section 4.1, Figure 1] The central claim that attention maps 'coherently align with object regions over time' and that this 'confirm[s] the interpretability of Wan's attention maps' is not supported by the presented evidence. Only a handful of prompts are shown, judged by eye, with no multiple seeds, no quantitative overlap between attention and object regions, and no negative cases. Section 5.2 concedes that maps were 'noisy, inconsistent, or visually unintelligible' for many tokens. Please either soften the claim to a qualitative demonstration, or add a systematic localization evaluation (e.g., thresholded attention vs. object segmentation/boxes across prompts, seeds, and frames), and report failure cases. This is load-bearing because the artistic and analytical uses rest on the alignment claim.
[Section 3, Visualization] The reshape from the flat latent embedding to the latent tensor [Frames, Height, Width] is the load-bearing operation: the claim that bright heatmap values correspond to object pixels in the generated video depends entirely on the ordering of flattened video latents and on the upscaling. The paper does not specify the flattening order, the interpolation method, or any validation of this correspondence. Please specify Wan's tensor layout and validate the correspondence concretely—for instance, by measuring whether the attention peak for the 'cat' token overlaps the cat's location in the generated frames (using a segmentation mask or bounding box) across several seeds and frames.
[Section 3, Appendix B] The tool is the main contribution but is not released, and no code link is provided; only the artistic video is public. The Appendix gives full settings for the multi-object probe, but the single-object and kiss probes in Section 4.1 do not report seed, resolution, frame count, or inference steps. For reproducibility, release the extraction/visualization code and provide complete configurations for all probes.

minor comments (4)

[Section 2, unnumbered equation] The equation softmax(QK^T) omits the standard 1/sqrt(d_k) scaling and the Value projection. Since the paper is about attention maps, it would help to state the full attention formula and clarify that the stored tensor corresponds to the softmax weights, not the attention output.
[Section 4.1] The text says 'additional results in Appendix A,' but the supplementary results are in Appendix B. Please correct the cross-reference.
[Figure 4 caption] The caption is confusing: it refers to a left-grid, right-grid, and far-left image without clearly specifying which panels correspond to the first frame, last frame, and the step-6 intermediate output. Please clarify the layout.
[Figure 3 caption] The three grids are described as first, middle, and final diffusion steps, but the caption does not label the subtitles within each grid. Adding explicit labels or a legend would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: attention maps are read out directly from the model; the interpretability claim rests on qualitative observation, not on a fitted input or self-citation chain.

full rationale

The paper does not derive a prediction from fitted parameters, and no load-bearing argument reduces to its own inputs. The extraction method intercepts cross-attention computations during generation and stores softmax(QK^T) maps (Section 3), so the visualized maps are direct model internals rather than quantities fitted to the outputs they are said to explain. The central claim that attention maps 'coherently align with object regions over time' (Section 4.1) is supported by selected examples and supplementary grids; it is a qualitative empirical assertion, not a definitional or statistical necessity. The paper explicitly concedes in Section 5.2 that maps can be 'noisy, inconsistent, or visually unintelligible,' which weakens generalizability but does not introduce circularity. The only self-citation is Network Bending [2], invoked in the conclusion as an aspirational future direction ('opens the door to network bending'); it is not used to justify the extraction method, the interpretability claim, or any uniqueness/ansatz assumption. No equation equates an output with an input, no fitted parameter is renamed as a prediction, and no self-citation carries the argument. Hence there is no significant circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper's own claims rest on standard attention math, the specific Wan architecture, and visual inspection. There are no fitted parameters or invented entities; the experimental settings (seed, guidance, prompts) are reporting choices, not fitted degrees of freedom.

assumptions (4)

domain assumption Cross-attention in Wan is computed as softmax(QK^T) with queries from text tokens and keys from video latents.
Section 2 states the attention equation and the mapping of tokens to video regions; the paper does not re-derive this, it depends on the Wan architecture.
domain assumption Visual alignment of attention maps with objects can be assessed by visual inspection.
Section 4.1 interprets heatmaps as aligning with objects without a quantitative metric; this is the core evidence for interpretability.
domain assumption The reshape from flat latent embedding to [Frames x Height x Width] preserves the spatial and temporal order needed for meaningful visualization.
Section 3 (Visualization) assumes the latent tensor can be reshaped to video dimensions, but no check of correspondence is provided.
domain assumption Wan2.1 1.3B is representative of video diffusion transformers.
The paper generalizes from a single model, Wan2.1 1.3B, without arguing representativeness.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Attention of a Kiss: Exploring Attention Maps in Video Diffusion for XAIxArts." pith.science (2026). https://pith.science/paper/I6JODFUP

@misc{pith2026250905323,
  author       = {Pith},
  title        = {Pith review of: Attention of a Kiss: Exploring Attention Maps in Video Diffusion for XAIxArts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I6JODFUP}},
  note         = {Machine review of arXiv:2509.05323}
}

read the original abstract

This paper presents an artistic and technical investigation into the attention mechanisms of video diffusion transformers. Inspired by early video artists who manipulated analog video signals to create new visual aesthetics, this study proposes a method for extracting and visualizing cross-attention maps in generative video models. Built on the open-source Wan model, our tool provides an interpretable window into the temporal and spatial behavior of attention in text-to-video generation. Through exploratory probes and an artistic case study, we examine the potential of attention maps as both analytical tools and raw artistic material. This work contributes to the growing field of Explainable AI for the Arts (XAIxArts), inviting artists to reclaim the inner workings of AI as a creative medium.

Figures

Figures reproduced from arXiv: 2509.05323 by the authors.

**Figure 2.** Attention of a Kiss: Video art study made from the attention maps of the token "kiss" in the open source video model Wan. 5 Discussion 5.1 Usefulness of Attention Maps for Artists Visualizing attention maps offers artists a valuable means of understanding how their textual prompts influence the visual outputs of generative video models. By revealing which regions of an image or video frame a given token attends to … view at source ↗

**Figure 1.** Comparison of attention maps from left-to-right [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 3.** Cross-attention maps for the token "cat" across all 30 transformer blocks. Grids from left to right correspond to the first, middle, and final diffusion steps. B.2 Attention in a Specific Transformer Block As noted above, attention often consolidates around object regions in the middle transformer blocks. In [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Cross-attention for the token "cat" in transformer block 15 across all diffusion steps. Left-grid corresponds to the first video frame; Right-grid corresponds to the last video frame. The far-left image is the intermediate output at diffusion step 6. While visually ind…

**Figure 5.** Figure 5: Cross-attention for the token "cat" across all 12 attention heads in block 15, visualized over 25 diffusion steps. Columns represent attention heads; rows represent diffusion steps [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Cross-attention map for "cat" averaged across all heads, blocks, and diffusion steps [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 2 canonical work pages

[1]

Samira Abnar and Willem Zuidema. 2020. Quantifying Attention Flow in Trans- formers. doi:10.48550/arXiv.2005.00928 arXiv:2005.00928 [cs]

arXiv doi:10.48550/arxiv.2005.00928 2020
[2]

Terence Broad, Frederic Fol Leymarie, and Mick Grierson. 2021. Network Bending: Expressive Manipulation of Deep Generative Models. doi:10.48550/arXiv.2005. 12420 arXiv:2005.12420 [cs]

work page Pith review arXiv doi:10.48550/arxiv.2005 2021
[3]

Nick Bryan-Kinns, Shuoyang Jasper Zheng, Francisco Castro, Makayla Lewis, Jia- Rey Chang, Gabriel Vigliensoni, Terence Broad, Michael Clemens, and Elizabeth Wilson. 2025. XAIxArts Manifesto: Explainable AI for the Arts. doi:10.1145/ 3706599.3716227 arXiv:2502.21220 [cs]

arXiv 2025
[4]

Hila Chefer, Shir Gur, and Lior Wolf. 2021. Transformer Interpretability Beyond Attention Visualization. doi:10.48550/arXiv.2012.09838 arXiv:2012.09838 [cs]

arXiv doi:10.48550/arxiv.2012.09838 2021
[5]

Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What Does BERT Look At? An Analysis of BERT’s Attention. doi:10.48550/arXiv. 1906.04341 arXiv:1906.04341 [cs]

arXiv doi:10.48550/arxiv.1906.04341 2019
[6]

Alec Helbling, Tuna Han Salih Meral, Ben Hoover, Pinar Yanardag, and Duen Horng Chau. 2025. ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features. doi:10.48550/arXiv.2502.04320 arXiv:2502.04320 [cs]

arXiv doi:10.48550/arxiv.2502.04320 2025
[7]

Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. 2022. Video Diffusion Models. doi:10.48550/arXiv. 2204.03458 arXiv:2204.03458 [cs]

arXiv doi:10.48550/arxiv.2204.03458 2022
[8]

Bingyan Liu, Chengyu Wang, Tongtong Su, Huan Ten, Jun Huang, Kailing Guo, and Kui Jia. 2025. Understanding Attention Mechanism in Video Diffusion Models. doi:10.48550/arXiv.2504.12027 arXiv:2504.12027 [cs]

arXiv doi:10.48550/arxiv.2504.12027 2025

Show all 16 references

[9]

2013.A History of Video Art

Chris Meigh-Andrews. 2013.A History of Video Art. A&C Black

2013
[10]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Models (Stable Diffusion). InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10684–10695

2022
[11]

Cynthia Rudin, Chaofan Chen, Zhi Chen, Haiyang Huang, Lesia Semenova, and Chudi Zhong. 2021. Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges. doi:10.48550/arXiv.2103.11251 arXiv:2103.11251 [cs]

2021 doi arXiv
[12]

Raphael Tang, Linqing Liu, Akshat Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. 2022. What the DAAM: Interpreting Stable Diffusion Using Cross Attention. doi:10.48550/arXiv.2210. 04885 arXiv:2210.04885 [cs]

2022 doi arXiv
[13]

Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. 2019. Towards Accurate Generative Models of Video: A New Metric & Challenges. doi:10.48550/arXiv.1812.01717 arXiv:1812.01717 [cs]

2019 doi arXiv
[14]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. InAdvances in Neural Information Processing Systems, Vol. 30. Curran Associates, Inc

2017
[15]

WanTeam, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Panden...

2025 doi arXiv
[16]

cinematic video of a cat playing with a soccer ball in front of the Eiffel Tower, realistic, 8k, high quality, masterpiece, best quality

Yuxin Wen, Jim Wu, Ajay Jain, Tom Goldstein, and Ashwinee Panda. 2025. Anal- ysis of Attention in Video Diffusion Transformers. doi:10.48550/arXiv.2504.10317 arXiv:2504.10317 [cs] ACM C&C XAIxArts 2025, June 23, 2025, Online Cole & Grierson A Media Files (1)Attention of a Kiss...

2025 doi arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Samira Abnar and Willem Zuidema. 2020. Quantifying Attention Flow in Trans- formers. doi:10.48550/arXiv.2005.00928 arXiv:2005.00928 [cs]

arXiv doi:10.48550/arxiv.2005.00928 2020

[2] [2]

Terence Broad, Frederic Fol Leymarie, and Mick Grierson. 2021. Network Bending: Expressive Manipulation of Deep Generative Models. doi:10.48550/arXiv.2005. 12420 arXiv:2005.12420 [cs]

work page Pith review arXiv doi:10.48550/arxiv.2005 2021

[3] [3]

Nick Bryan-Kinns, Shuoyang Jasper Zheng, Francisco Castro, Makayla Lewis, Jia- Rey Chang, Gabriel Vigliensoni, Terence Broad, Michael Clemens, and Elizabeth Wilson. 2025. XAIxArts Manifesto: Explainable AI for the Arts. doi:10.1145/ 3706599.3716227 arXiv:2502.21220 [cs]

arXiv 2025

[4] [4]

Hila Chefer, Shir Gur, and Lior Wolf. 2021. Transformer Interpretability Beyond Attention Visualization. doi:10.48550/arXiv.2012.09838 arXiv:2012.09838 [cs]

arXiv doi:10.48550/arxiv.2012.09838 2021

[5] [5]

Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What Does BERT Look At? An Analysis of BERT’s Attention. doi:10.48550/arXiv. 1906.04341 arXiv:1906.04341 [cs]

arXiv doi:10.48550/arxiv.1906.04341 2019

[6] [6]

Alec Helbling, Tuna Han Salih Meral, Ben Hoover, Pinar Yanardag, and Duen Horng Chau. 2025. ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features. doi:10.48550/arXiv.2502.04320 arXiv:2502.04320 [cs]

arXiv doi:10.48550/arxiv.2502.04320 2025

[7] [7]

Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. 2022. Video Diffusion Models. doi:10.48550/arXiv. 2204.03458 arXiv:2204.03458 [cs]

arXiv doi:10.48550/arxiv.2204.03458 2022

[8] [8]

Bingyan Liu, Chengyu Wang, Tongtong Su, Huan Ten, Jun Huang, Kailing Guo, and Kui Jia. 2025. Understanding Attention Mechanism in Video Diffusion Models. doi:10.48550/arXiv.2504.12027 arXiv:2504.12027 [cs]

arXiv doi:10.48550/arxiv.2504.12027 2025