Stable-Video-Infinity

Why we included this project

Generating video that stays coherent past a few seconds is one of the harder problems in diffusion-based video generation, and Stable Video Infinity is built around solving it. Instead of treating each stretch as a fresh generation problem, it recycles and corrects its own output as it goes, so long scenes hold together rather than drifting into artifacts. From the VITA lab at EPFL and accepted as an ICLR 26 oral, it lets you chain text prompts into a single narrative with each clip controlled separately, and the same model family covers talking faces and dance generation. That makes it a solid pick for researchers and for content teams who want prompt-driven long-form footage without stitching together independent clips.