Spent the whole night finally getting Wan2.2’s text-to-video workflow running locally in ComfyUI. Gonna jot this down so nobody else has to bang their head against the wall.
It’s an open-source text-to-video model using the MoE (Mixture of Experts) architecture. Official word is it’s smoother than 2.1 with fewer artifacts, and can do 720p. After actually using it, two things I gotta stress:
First, you must slap on that lightning x2v LoRA. With it, you can only set the sampling steps to 4—anything more and it’ll just error out. Without it, a single generation takes me nearly half an hour; with it, it’s orders of magnitude faster.
Second, VRAM is tight as hell. My 24G card barely handles 480p at 81 frames—anything higher and it’s OOM city. If you want 720p, just bite the bullet and get a bigger card. Those red nodes mean missing custom nodes—just hit “Install Missing Nodes” in the Manager.
You’ll need to download six models total: two main models (high-noise and low-noise), plus VAE, text encoder, and two LoRAs. Don’t miss any. First-time setup is a bit of a maze, but once it’s running, it’s actually pretty stable.