Wan 2.2

Wan 2.2 by Alibaba: the video model of most lessons in the app; the 14B version works as two experts, the 5B version as one small model.

#What it is

Wan 2.2 (2025) improves motion and look over Wan 2.1. The 14B models (t2v and i2v) are a pair: a high_noise expert lays out the clip and a low_noise expert finishes it, in two samplers (see Wan 2.2 two passes). The 5B model (ti2v) does both text and image to video in one model on cards with about 8 GB, with its own VAE. Further models: Animate (pose and face from a video), S2V (sound to video), Fun Control and InP, and VACE modules for 2.2.

#Files and nodes

• 14B: two Load Diffusion Model nodes (wan2.2_t2v_high_noise_14B and ..._low_noise_14B, or the i2v pair); Load CLIP with umt5_xxl (type wan); Load VAE with wan_2.1_vae (the 14B models use the 2.1 VAE); two ModelSamplingSD3; EmptyHunyuanLatentVideo or WanImageToVideo; two KSampler (Advanced).
• 5B: one Load Diffusion Model (wan2.2_ti2v_5B), the same text encoder, wan2.2_vae, Wan22ImageToVideoLatent, one sampler.

#Settings that work

14B: 832×480 or 1280×720, 81 frames at 16 fps, 20 steps handed over at 10, cfg 3.5, shift 8; with lightx2v LoRAs 4 steps handed over at 2, cfg 1. 5B: 1280×704, 121 frames at 24 fps. Prompt: detailed sentences with action and camera.

#What fits

LoRAs for Wan 2.2 and the right expert (high or low files); many Wan 2.1 14B LoRAs work too. The 5B and 14B add-ons are not interchangeable.

#How the app helps

Shown as Wan 2.2 on model cards. Most lessons of the groups Video, Video: the anchor picture and WanVideoWrapper are built on it, the 5B model among them in Video on a modest graphics card: Wan 2.2 5B.

#See also

• Wan 2.2 two passes
• Wan 2.1
• Speed-up LoRAs
• Lesson: Video from text