Stable Video Infinity: AI tool for long, coherent video generation

Researchers at the Swiss Federal Institute of Technology Lausanne (EPFL) have unveiled a new AI-based tool that addresses a key challenge in video generation: the loss of temporal coherence. The system, called Stable Video Infinity (SVI), has already caught the attention of the tech community.

Most current video generation models can only produce short clips lasting from a few seconds to about half a minute. When the duration increases, images start to distort: characters change, scenes become unstable, and the logic of the sequence breaks down. Known as drift, this effect has long been considered almost unavoidable.

The Visual Intelligence for Transportation (VITA) lab team proposed an unconventional approach to training models. Instead of ignoring errors that arise during video generation, their new method—"recurrent error retraining"—deliberately reintroduces these glitches into the learning process. In practice, this means the AI learns to handle its own distortions.

Project lead Professor Alexander Alahi compares the approach to training a pilot in severe turbulence. Learning from errors makes the system more robust and allows it to maintain stability even during extended generation. This principle underpins Stable Video Infinity. Unlike existing solutions, which often degrade after 20–30 seconds, SVI can create coherent and detailed videos lasting several minutes or longer.

The team also introduced the LayerSync method, which helps the AI synchronize internal logic when working with video, images, and audio. Together, these technologies pave the way for more reliable autonomous systems and large-scale generative media. The project is already open-source on GitHub, where it has garnered over 2,000 stars, and the research was presented at the ICLR 2026 conference, highlighting its significance for the future of generative technologies.