ComfyUI setup: Wan 2.2, InfiniteTalk, SCAIL and LTX-2

This tutorial covers ComfyUI video generation on local hardware, with model downloads, folder placement and workflow setup. The speaker tested everything with 24 GB of VRAM and says some workflows can run with 16 GB. Certain Wan 2.2 GGUF setups may fit 8 GB, but that does not apply to every workflow shown.

Wan 2.2 setup uses matching high-noise and low-noise models, a text encoder and a VAE. The demonstrated image-to-video workflow produces one to five seconds at 16 fps, with HD as its stated resolution limit. The tutorial also explains prompt writing and LoRA connections. Wan Animate transfers motion from a reference video and can replace the character or the background too. For SCAIL, the speaker recommends human motion references after encountering pose-detection failures with cartoons; multiple-character tests worked best in portrait 9:16.

InfiniteTalk combines a portrait with supplied speech and extends clips in three-second stages. The prepared workflows reach 34 seconds, though the speaker reports generation times above half an hour for longer clips. LTX-2 covers generated and custom audio, with dimensions divisible by 32 and frame counts following an eight-plus-one rule. The speaker found its results less consistent than Wan's.

The closing sections cover SeedVR2 upscaling, frame interpolation and RunningHub cloud workflows. Sage Attention is optional; the speaker reports GPU compatibility problems in their Windows setup.