Descriptions:
Veteran AI presents two ComfyUI workflows for upscaling existing videos using LTX-2, demonstrating that the model’s native upscaling capability can be applied to externally sourced footage—not just clips the model generated itself. Both workflows are adapted from LTX-2’s image-to-video pipeline, with the generation stage stripped out and replaced by an encode-upscale-decode pass.
The simplified workflow processes video at 25fps (LTX-2’s recommended frame rate), extracts 121 frames, resizes to 1280×720, and encodes to latent space at half resolution before upscaling latents back to full size. The LCM sampler runs at 4 steps with CFG 1.0. A notable technical hurdle is LTX-2’s mandatory audio input: since upscaling a silent video has no audio, the solution is to generate an empty audio latent as a placeholder and combine it with the video latent before sampling. The advanced workflow improves on this by substituting a separately prepared high-quality image for the auto-extracted first frame, which serves as the guidance condition.
A controlled bypass experiment proves the first-frame condition is essential: disabling the guidance node causes complete character consistency failure and scene degradation. Both workflows use Kijai’s reorganized node structure with separate audio and video VAE loaders for clarity. All workflows are available on RunningHub’s ComfyUI platform for direct testing and modification.
📺 Source: Veteran AI · Published January 13, 2026
🏷️ Format: Tutorial Demo
