Explainability-in-Action: Enabling Expressive Manipulation and Tacit Understanding by Bending Diffusion Models in ComfyUI

Explainability can serve creative purposes beyond just transparency: The paper demonstrates that making AI systems' internal workings visible and manipulable can help artists develop intuition about how generative models work, moving beyond traditional "black box" explanations toward active engagement and understanding through hands-on experimentation.

Interactive manipulation builds tacit understanding: By allowing artists to directly modify different components of diffusion models through ComfyUI's interface, users develop practical knowledge about how each part influences outputs—similar to how craftspeople learn through repeated, reflective practice rather than abstract instruction.

Large-scale models can be made more transparent without retraining: Rather than requiring artists to train their own smaller models, the researchers show that even large text-to-image diffusion systems can be "bent" and inspected through accessible interfaces, democratizing control over generative AI tools.

Craft-based approaches to AI explainability matter: The paper argues that explainability in creative contexts should emphasize long-term, hands-on engagement and "learning by doing" rather than just providing technical documentation, making AI systems more understandable through active experimentation.

Tool design shapes AI agency and understanding: The integration of model-bending capabilities into ComfyUI's node-based interface demonstrates how thoughtful interface design can give users greater agency over generative systems and support deeper comprehension of how AI components work together.