Pulse of Motion: Teaching AI videos to obey the laws of physics

Modern AI video generators have gotten scarily good at rendering photorealistic textures, but something almost always feels off the moment things start moving.

A falling object drifts like it is floating in deep space, or a person walks with an unnatural, caffeinated twitch. On March 15, researchers from Texas A&M University addressed this exact headache by open-sourcing Pulse of Motion, a clever sub-2GB tool designed to correct what they call "chronometric hallucinations."

The missing internal clock in video models

The issue is not that today's frontier models lack visual intelligence. The problem is that they lack a reliable internal clock.

During training, billions of video clips filmed at wildly different capture rates are shoved into a model, and most pipelines simply normalize everything into a standard container rate like 24 or 16 frames per second. The generator learns how a humming bird looks, but it has no real understanding of how fast gravity pulls or how rapid wings beat in the physical world. It is essentially guessing playback velocity on vibes alone.

Texas A&M's team found that even massive general-purpose vision-language models make terrible judges of temporal physics. They simply cannot tell whether a clip is moving at realistic real-world speeds.

Fixing gravity with a sub-2GB motion predictor

Instead of trying to retrain an entire multi-billion-parameter generator from scratch, Pulse of Motion tackles the problem from the outside.

The sub-2GB model inspects the raw visual dynamics of a video and calculates its true "Physical FPS" (PhyFPS). Once it estimates how fast that real-world action should actually unfold, it retimes the footage accordingly.

The results speak for themselves across several public demos. A clip of a snake slithering across polished wood was nominally 24 fps, but Pulse of Motion determined its true physical motion matched 60.2 fps, turning an awkward float into a snappy glide. Onion rings frying in hot oil jumped from 24 fps to 58.9 fps, while fish struggling inside a net was slowed down from 24 fps to a much more realistic 15.0 fps.

The team has released the code, download instructions for consumer hardware, and a public leaderboard to benchmark how well different generators adhere to real-world physics.

Why it matters

We think Pulse of Motion is a brilliant example of lightweight post-processing outsmarting brute-force scale.

Right now, AI video labs are burning millions of dollars trying to make next-generation foundation models inherently understand physics. Yet temporal pacing remains a persistent blind spot across the industry. Texas A&M proved that you do not need a cluster of enterprise GPUs to fix movement pacing; an accessible tool running locally on an everyday laptop can restore visual credibility in seconds.

For digital creators and VFX editors, this eliminates hours of manual speed-ramping and trial-and-error prompting. If a generated clip looks floaty, you can run it through Pulse of Motion to lock down realistic inertia.

The limitation, of course, is that retiming cannot magically invent missing frames if the base generator failed to draw complex motion in the first place. But by isolating temporal fidelity as a standalone metric, this work draws a clear line: visual fidelity alone is no longer enough if your AI cannot keep time.