Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms
Yueyan Li, Haibo Wang, Caixia Yuan, Xiaojie Wang
Beijing University of Posts and Telecommunications
Published September 20, 2026 · on the feed September 22, 2026 · ▲ 25 · arXiv 2609.23658
The abstract
Despite impressive visual quality, state-of-the-art video diffusion models often generate content that violates real-world physical laws. While existing solutions rely on external priors or specialized data, we investigate the root cause by exploring the internal mechanisms of these models. Specifically, we present the first interpretability study on the ''motion planning'' process of text-to-video diffusion models, revealing how motion trajectories form during early denoising stages. Building upon the ''first shape, then details'' finding, we combine cross-attention trajectory patterns with causal head contributions to identify a specific subset of attention heads driving motion planning. Further, our self-attention analysis shows that Rotary Position Embedding (RoPE) induces excessive spatial attention decay. This causes early candidate regions to prematurely lock into physically implausible positions, suppressing reasonable trajectories in adjacent frames and triggering generation failure modes. To address this fundamental flaw, we propose a lightweight architectural modification that scales the frequency of RoPE across different denoising steps. This strategy reduces excessive attention decay, helping the model explore better candidate regions to establish coherent physical motion. Finally, training-free and training-based experiments confirm the effectiveness of our approach in enhancing the physical commonsense of generated videos.
The Brief
2:27A short podcast summary of the paper. Written and voiced by AI.
Transcript
Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms. By Yueyan Li and three others, from the Beijing University of Posts and Telecommunications. Published on the twentieth of September. Its code repository is worked out: about one thousand source files, with a readme.
State-of-the-art video diffusion models produce impressive visuals that often break real-world physical laws. Existing fixes bring in external priors or specialized data. This work looks inward instead, for the root cause, and places it in the attention mechanism. The claim is that Rotary Position Embedding, RoPE, induces excessive spatial attention decay, and that this decay locks motion into physically implausible paths early in generation.
The route to that claim begins with what the authors present as the first interpretability study of motion planning in text-to-video diffusion models. It shows that motion trajectories form during the early denoising stages: first shape, then details. Building on that, they combine trajectory patterns in cross-attention with the causal contributions of individual heads, and pick out a specific subset of attention heads that drive motion planning.
The self-attention analysis then supplies the failure. The spatial attention decay induced by RoPE is excessive. Early candidate regions prematurely lock into positions that are physically implausible. Reasonable trajectories in adjacent frames are suppressed as a result, and that is what triggers the failure modes in generation.
A fix follows from the diagnosis: a lightweight architectural change that scales the frequency of RoPE differently across denoising steps. That reduces the excessive decay, letting the model explore better candidate regions and establish coherent physical motion. Experiments that are training-free, and experiments that involve training, both confirm that it improves the physical commonsense of the generated videos.
The study concerns text-to-video diffusion models, and specifically the motion-planning phase in early denoising. The mechanism identified is one cause located inside attention, through RoPE's spatial decay, and a particular subset of heads. The remedy is judged on physical commonsense in generated video, and it is shown to work both without training and with it.
This finding relocates the problem. Instead of treating physics violations as a gap in knowledge to be filled from outside, it traces them to a timing problem: candidate positions settling too soon, because attention decays too steeply across space. Adjusting how RoPE behaves across denoising steps is presented as enough to loosen that early lock.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a beautiful video where the ball moves wrong.
You generate videos from text with a state-of-the-art model. They look stunning, but a thrown ball swerves, or an object jumps to an impossible spot.
The road ahead
- 1 · The problem ~6 min Video models make beautiful clips that break physics, and the usual fixes patch the output without explaining why the model fails.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Motion is decided in the first denoising steps, where rotary position embedding makes attention fade too fast with distance; weakening it early lets better trajectories win.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Find when motion forms, find which heads plan it, find the self-attention flaw, then rescale RoPE frequency in the early steps.
- 6 · The experiment ~6 min You’ll guess: Where did the authors locate a root cause of physics violations in text-to-video diffusion models?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.