Generated video has a blooper reel nobody asked for: a basketball drops from mid-air, bounces a few times, then freezes in place — and with the same prompt across four random seeds, the failure modes differ each time. These aren't cherry-picked fails; they're the everyday output of state-of-the-art video diffusion models. Instead of stuffing physics priors or curated data into the model, a team from BUPT's Center of Intelligent Science and Technology opened the model up and asked a more basic question: at which step, and by which mechanism, is a motion trajectory actually decided?
Motion planning: the first 5 steps decide everything
Working with Wan2.1-T2V-1.3B, the team found that "motion planning" — the process that lands text semantics onto object positions frame by frame — happens largely within the first 5 of 50 denoising steps, following a "first shape, then details" pattern. Combining cross-attention trajectory patterns with causal intervention experiments, they separated the attention heads that genuinely drive motion from those that merely display its pattern.
The culprit: RoPE's spatial anchoring bias
Following that thread into self-attention exposed the root cause: 3D RoPE (Rotary Position Embedding) induces a spatial anchoring bias — a query tends to attend to spatially nearby coordinates in other frames, so a physically plausible candidate farther away receives far too little attention. Worse is the "early commitment" effect: candidate regions initially compete, but once a few frames settle early, spatial anchoring rallies support for nearby positions in other frames even when they violate physics. In the paper's most emblematic failing seed, frame 14 switches from a lower candidate to a higher one after frame 10 stabilizes — and the basketball simply hangs in mid-air.
The fix is light: rescale RoPE frequency in the first 5 steps
With the root cause diagnosed, the fix turns out to be cheap. Training-free: scale down the height and width RoPE frequencies during only the first 5 denoising steps to weaken the excessive spatial decay, then restore the originals — no weight updates required. The training-based variant adapts self- and cross-attention with LoRA and uses λ = 0.70 for the first 5 steps at inference. On 344 VideoPhy cases judged by humans for semantic adherence and physical commonsense, the largest gains came on interactions involving solids. The paper bills itself as the first interpretability study of the motion-planning process in text-to-video diffusion models; code and project page are released alongside (https://arxiv.org/abs/2609.23658, https://siriuslala.github.io/physics).
So what
Days ago the conversation was still about how to score video models' physics; now someone has produced the pathology slide. The counterintuitive takeaway worth chewing on: a physics problem didn't need a physics solution — adjusting the frequency of a position encoding stopped the bleeding. Some of what looks like "not understanding the world" in video models may just be attention budget allocated too early and too rigidly in space. Next time an object levitates, don't blame the data first — check the RoPE.