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 a series of artistic case studies, 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 & tables
Figure 1: Intercepting the signal. Artistic study using attention maps from the internal diffusion process as creative material rather than the final video outputs. This diptych shows the structural similarities in the attention maps between an intimate kiss and a cosmic collision, revealing how the neural network sees impact across different scales.
Figure 2: Visualizing Attention . A cross-attention map for the token “cat” overlaid on the original generated video. The heatmap visualizes the spatial correspondence between text and image: warmer colors (yellow/orange) indicate regions where the token “cat” has the strongest influence on the generated pixels, while cooler colors (blue) indicate areas of low association.
Figure 3: Visualization Interface . Web-based UI enabling both macro-level and granular exploration of cross-attention maps across diffusion steps, transformer layers, and individual attention heads. The default view shown here displays per-token attention maps averaged across each diffusion step: from left to right, the sequence reveals the development of attention from unfocused to concrete structure.
Probe Type
Prompt Text (Tracked Tokens in Bold)
Figure
Single Object
A cute black cat playing in a green field
Fig. 2
Multiple Objects
A cute black cat playing with a ball in front of the Eiffel Tower
Fig. 4, 6
Motion
A red car driving fast through the rocky mountains
Fig. 5
Abstract Action
A cinematic romantic kiss between two people
Fig. 5
Table 1: Experimental Dataset. A list of the exploratory probes used to validate the tool. Tracked tokens are highlighted in bold within the prompt text.
Figure 4: Semantic separation and relational context. Attention maps for “cat” (left), “ball” (center), and “Eiffel Tower” (right). While the background architecture is clearly separated (right), the maps for the interacting subjects (cat and ball) exhibit significant overlap. This suggests the attention mechanism captures the relational context of the “playing” action rather than strictly isolating object contours.
Figure 5: Motion and Abstraction. Left: Attention maps for the token “car” at two different timestamps. The model successfully tracks the object’s movement across the frame despite changes in position. Right: A comparison of the concrete token “kiss” versus the abstract token “romantic” within the same frame. While “kiss” localizes to the physical interaction, “romantic” appears as a diffuse, scene-wide influence.
Figure 6: A comparison of the attention map (left) versus the decoded video content (right) at diffusion step 2 of 15. While the video output appears as unrecognizable noise, the attention map already displays strong compositional coherence, anticipating the final structure.
Figure 7: Left: Attention maps averaged by Layer, showing the evolution of attention across model depth. Right: Attention maps by Head, revealing the granular diversity of individual attention mechanisms for a single token.
Figure 8: Attention of a Kiss : A video art study constructed from the attention maps of the token “kiss” in the open source video model Wan. The sequence illustrates the transition from latent noise to semantic structure.
Figure 9: Multi-Head Attention : A visualization of the token “cinematic” across eight different generations. The abstraction of the attention map obscures identity markers, reducing the scene to the gestural essence of the interaction.
Figure 10: Diffusion Collision : A diptych comparing the attention dynamics of a human kiss (left) and a planetary collision (right). The study highlights the morphological similarities in how the model constructs “impact” across vastly different semantic scales.
We present AttentionBender, a tool that manipulates cross-attention in Video Diffusion Transformers to help artists probe the internal mechanics of black-box video generation. While generative outputs are increasingly realistic, prompt-only control limits artists' ability to build intuition for the model's material process or to work beyond its default tendencies. Using an autobiographical research-through-design approach, we built on Network Bending to design AttentionBender, which applies 2D transforms (rotation, scaling, translation, etc.) to cross-attention maps to modulate generation. We assess AttentionBender by visualizing 4,500+ video generations across prompts, operations, and layer targets. Our results suggest that cross-attention is highly entangled: targeted manipulations often resist clean, localized control, producing distributed distortions and glitch aesthetics over linear edits. AttentionBender contributes a tool that functions both as an Explainable AI style probe of transformer attention mechanisms, and as a creative technique for producing novel aesthetics beyond the model's learned representational space.
Generative video models have achieved remarkable visual fidelity, yet their prompt-only interface offers thin creative agency and obscures the model's material process from the artists working with it. We present Real-Time AttentionBender, a tool that extends the practice of network bending across the full depth of the video diffusion transformer (DiT) and brings it into live, interactive generation. Built as a plugin within the DayDream Scope ecosystem and wrapping open-source real-time Wan pipelines, the tool exposes self-attention, cross-attention, and the feed-forward network as independently manipulable surfaces, with targeting down to individual diffusion steps, DiT layers, prompt tokens, and hidden neurons. The immediacy of live manipulation affords what we call "material intimacy" with the model: a responsive, near-mechanistic feel for how specific layers and neurons shape generated video. We position the tool as simultaneously an XAIxArts probe into transformer internals and an expressive instrument for discovering aesthetics outside the model's default representational space.
Adam Cole, Rebecca Fiebrink, Mick Grierson
Creative Computing Institute University of the Arts London London, UK
Diffusion-based text-to-image models can synthesize complex and highly structured visual content, yet the emergence and evolution of semantic structure remain difficult to interpret. Many existing workflows rely on aggregated attention or scalar summaries that separate temporal change from image-space evidence. To address this gap, we present a visual analytics framework for exploring attention dynamics in diffusion models: the step-indexed evolution of token-level cross-attention maps, their temporal concentration, and their spatial relationships. Our approach enables structured analysis of attention behavior across generation steps by integrating quantitative measures with data-driven stage identification in an interactive workflow. Case studies on a structured 60-prompt Stable-Diffusion-class benchmark illustrate recurring, interpretable patterns within this setting and show how linked temporal and spatial views facilitate the observation and discussion of generative processes, supporting more effective human-AI collaboration.