Researchers at EPFL have created a new way to fix drift in generative video. Drift is the problem that makes AI videos become messy and illogical after only a few seconds. Small errors such as blurred faces or wrong shapes get bigger with each new frame. Most AI models learn only from perfect pictures so they fail when they meet their own mistakes in real use.
The new method is called retraining by error recycling. It works by letting the AI make a video then measuring the errors between the real output and the perfect image. These errors are saved and fed back into the model during training. The AI now learns to correct itself and keep images clear and logical even when the starting frame is imperfect. After this training the model stays stable for much longer.
How the error recycling process works
The approach needs little extra computing power or data. It has been built into a system named Stable Video Infinity or SVI. SVI can now produce good videos that last several minutes or more. The code is open source and has been tested against other AI systems. It has gained wide interest with over 1.9k stars on GitHub and strong community feedback. It will be presented at the 2026 International Conference on Learning Representations in April.
Researcher Wuyang Li explains that "unlike humans, generative AI rarely knows how to recover from its mistakes, which leads to drift. So we teach the models how to do this and how to remain stable despite imperfection.”
Researchers at EPFL have also used the same idea to make LayerSync. This method recycles not only visible errors but also internal signals from smarter parts of the model. It helps the AI improve faster and create better content across video images and sound without needing extra data. The work will be shown at the International Conference on Learning Representations in April. It will help build safer autonomous systems that interact better with people.