Chinese Researchers Found a Way to Make AI Video 3x Faster

AI video tools like Sora and Kling don't make a whole clip at once. They build it frame by frame, starting with random static and cleaning it up step by step until it looks real. The tricky part is keeping things consistent: if a character wears a red hat in one frame, they should still be wearing it 20 seconds later. The usual fix is to only move forward once the previous frame is completely finished — accurate, but slow.

A new paper from the University of Chinese Academy of Sciences questions that rule. Hongchen Zhang tested three ways of feeding past frames back into the model: one using only fully finished frames, and two that allow half-finished or progressively built frames to be reused. On VBench, a standard test for AI video quality, the two looser methods scored 85.24 and 85.60, beating the careful approach's 84.45. In longer videos, they also kept characters and motion more stable.

The bigger payoff is speed. Because you no longer have to wait for one frame to finish before touching the next, several frames can be cleaned up at the same time. That produced 1.57 to 2.83 times speedups in the paper's testing. The author also found a cheaper way to tune the model for quality: a small add-on patch, known as LoRA (a lightweight add-on that trains a big AI model cheaply), reached better results using just 2.15% of the adjustable settings a full retune would need.

The catch: this is a research paper, not a product. The numbers come from the author's own setup, so real-world gains may be smaller. Still, it points to a near future where AI video is faster, cheaper, and stable enough for longer clips — the groundwork for live, on-the-fly AI video.