AttentionBender gives artists real-time neuron-level control over video AI

Generative video models have reached stunning visual quality, but their prompt-only interface limits creative agency—artists can only describe what they want, not reshape how the model thinks. Real-Time AttentionBender shatters that wall. Built as a plugin within the DayDream Scope ecosystem and wrapping open-source real-time Wan pipelines, this tool lets users directly manipulate the transformer's internals: self-attention, cross-attention, and feed-forward networks. Targeting can be as granular as individual diffusion steps, DiT layers, specific prompt tokens, or even hidden neurons. The result is a live, interactive experience where artists can sculpt video generation as it happens, feeling exactly how each layer and neuron contributes to the final output—what the authors call “material intimacy.”

Beyond creative expression, the tool serves as an XAIxArts probe, letting both artists and researchers explore transformer internals in real time. It reveals how attention patterns and feed-forward activations shape motion, texture, and composition, offering a window into the model's default representational space—and a way to break out of it. This positions AttentionBender as both an expressive instrument for discovering novel aesthetics and a practical method for understanding video diffusion transformers. By bridging art and explainable AI, it opens the door to a new era of hands-on, mechanistic interaction with generative models.