ⓘ Gemini
If you’ve ever used video generation models, one thing is consistent across the board: That means you’re limited to short clips, usually between 5 and 20 seconds. The reason this limit exists is because of something called “drift.” Drift causes scene and character characteristics to be lost from frame to frame, causing the output to become inconsistent over time.
To address this problem, researchers at EPFL’s Visual Intelligence for Transportation (VITA) laboratory have developed a new training method called “retraining with error recycling.” This approach intentionally feeds defects and deformations that naturally occur during generation into the model, rather than discarding them.
Professor Aleksandr Alahi likens the process to “training pilots in rough weather instead of clear blue skies.” By learning from its mistakes, AI becomes robust enough to stabilize itself when errors inevitably occur, rather than spiraling into randomness.
This method powers the new Stable Video Infinity (SVI) system. Unlike current models, which often fall apart after 30 seconds, SVI can produce consistent, high-quality video that lasts several minutes or more. The system is already making waves in the tech community. The open source code on GitHub has over 2,000 stars, and the research has been accepted for presentation at the 2026 International Conference on Learning Representations (ICLR).
The team is also debuting LayerSync, a related method that allows AI to modify internal logic across video, image, and sound generation. By combining these tools, we hope to design better autonomous systems and unlock the potential of truly long-form generative media.
Throughout my life, I have always been fascinated by and even addicted to technology and digital devices. I’m always amazed at the complexity of the simplest digital devices and systems around us. I’ve been writing and publishing online for about six years, but just over a year ago I found myself falling in love with the wonders of the smartphones and laptops we have in our hands every day. I developed a passion for learning about new devices and the technology that goes with them, and at some point I asked myself, “Why not get involved in writing about technology?” It’s no use saying that I followed up on that idea – it’s obvious. I am an open-minded person who finds endless joy in researching and discovering new information. I believe there is a lot to learn and life is very short, so I use my time wisely to learn new things. I’m a “bookworm” of the internet and digital devices. When I’m not writing, you can still find me on your devices. I explore and admire the beauty of nature and living things. I am a quick learner, adapt quickly to change, and always look forward to new adventures.
