Interpretability Study Links Attention Heads to Video Physics Violations
September 19, 2026
An interpretability study reveals that video diffusion models' motion planning occurs during early denoising stages, driven by specific cross-attention trajectory patterns. The research identifies that Rotary Position Embedding (RoPE) and certain attention heads contribute to physical inconsistencies in generated video.
HOW THIS AFFECTS YOU
●
researcherYou can target specific attention heads or embedding schemes to improve the physical realism of video generation.