Descriptions:
Veteran AI introduces SVI Pro, the updated release of the Smooth Video Interpolation model built on Wan 2.2, specifically designed to fix the “action loop” bug that plagued the original SVI—where generated characters would repeatedly return to their starting position instead of completing the intended motion.
The video covers three distinct ComfyUI workflows in detail. The first extends an existing video clip using SVI Pro’s latent-space approach. The second starts from a single still image, generates an initial video segment, and then sequentially extends it. The third enables theoretically infinite video extension through a looping pipeline. Central to all three is the “Anchor Frame” concept—the reference image VAE-encoded into latent space and carried through every extension step to maintain visual continuity—alongside the “Last Latent,” which encodes the final frame of each prior segment to guide subsequent motion.
The host explains why the Light X to V acceleration LoRA is mandatory (the Lightning variant causes instability), specifies which model versions to use for high-noise and low-noise sampling, and details the KJ Nodes extension by Kijai required for the SVI-specific latent node. The “Motion Latent Count” setting and video overlap stitching parameter are called out as particularly important tuning knobs. All workflows are available on RunningHub for direct testing.
📺 Source: Veteran AI · Published December 31, 2025
🏷️ Format: Tutorial Demo
