Video diffusion models can produce strikingly realistic footage, yet they frequently generate motion that contradicts basic physical laws. A new paper posted on Hugging Face investigates this problem not by adding external constraints or specialized training data, but by looking inside the models themselves.
The authors describe their work as the first interpretability study of the "motion planning" process in text-to-video diffusion models, showing how motion trajectories take shape during the early denoising stages. Building on the observation that generation follows a "first shape, then details" pattern, they combine cross-attention trajectory patterns with causal head contributions to isolate a specific subset of attention heads responsible for motion planning.
Their self-attention analysis points to a concrete culprit: Rotary Position Embedding (RoPE) induces excessive spatial attention decay. As a result, early candidate regions become prematurely locked into physically implausible positions, suppressing reasonable trajectories in adjacent frames and triggering generation failure modes.
To address the flaw, the researchers propose a lightweight architectural modification that scales the frequency of RoPE across different denoising steps. This reduces the excessive attention decay and helps the model explore better candidate regions, allowing it to establish coherent physical motion.
The paper reports that both training-free and training-based experiments confirm the approach improves the physical commonsense of generated videos, offering a fix that avoids heavy training overhead.