Stable Video Diffusion XT

Where the first Stable Video Diffusion checkpoint stopped at 14 frames, this XT release was fine-tuned to output 25 frames at 576 by 1024 from one conditioning image. The Hub repo holds two weight files of about 9.5 GB each: the default uses a VAE decoder retuned for frame-to-frame consistency, the alternative the standard per-frame image decoder. Stability measured around three minutes per clip on an A100 80GB at launch defaults; in Diffusers, offloading and small decode chunks bring memory toward consumer cards at the cost of speed and some flicker. There is no text prompt at all, so results depend on the source image plus motion and frame-rate hints, and the model card warns about weak faces, unreadable text, and occasional near-static output. Stability's reference scripts add an invisible watermark to outputs by default. A later 1.1 revision sits behind a click-through gate on the Hub. Weights fall under the Stability AI Community License, free for research and personal work and open to registered commercial users whose organizations earn under 1 million US dollars a year. Suited to image-to-video prototyping.