AttentionBender: Manipulating Cross-Attention in Video Diffusion Transformers as a Creative Probe

AI systems resist simple, predictable control: The paper found that manipulating specific attention mechanisms in video generation models produced "distributed distortions" rather than localized changes, suggesting that even when researchers can access a model's internal components, the system's behavior remains difficult to predict or control precisely.

Transparency tools can serve dual purposes for responsible AI: AttentionBender functioned both as an "Explainable AI style probe" and creative tool, demonstrating that technical interventions designed to understand AI systems can simultaneously enable new capabilities—highlighting the importance of considering multiple stakeholder needs when building interpretability tools.

Cross-attention mechanisms are highly entangled: The analysis of 4,500+ video generations revealed that different parts of the model's attention system are deeply interconnected, meaning changes intended to affect one aspect of generation often have unexpected ripple effects elsewhere—a finding relevant to understanding AI system complexity and potential failure modes.

Probing black-box systems requires hands-on experimentation: The researchers used "research-through-design" methodology to build intuition about how video diffusion models work, suggesting that meaningful AI transparency may require iterative, exploratory approaches rather than relying solely on theoretical analysis or documentation from model creators.