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

Attention maps can make AI video generation more transparent: The paper demonstrates that by visualizing cross-attention maps in video diffusion models, researchers can see how the AI system connects text prompts to specific visual elements across frames, providing a concrete way to understand what the model "pays attention to" during generation.

This transparency tool has dual value for artists and researchers: Rather than treating attention visualization as purely technical, the authors position it as both an analytical tool for understanding AI behavior and as raw creative material that artists can manipulate—similar to how early video artists worked with analog signals—suggesting new collaborative possibilities between AI systems and human creativity.

Current AI video models show interpretable temporal and spatial patterns: The study reveals that attention mechanisms in video generation exhibit consistent behaviors across time and space, meaning these patterns aren't random or completely opaque, which is important for assessing whether AI systems are making decisions we can actually understand and potentially predict.

Explainable AI extends beyond traditional domains: By framing this work as "XAIxArts" (Explainable AI for the Arts), the paper argues that responsible AI development shouldn't focus only on high-stakes domains like healthcare or criminal justice, but should also address creative AI systems that increasingly influence cultural production and artistic expression.