Wan 2.2: the two passes
Why the large Wan 2.2 models come as a pair, and how two samplers share one generation.
In plain words
Wan 2.2 14B is a "mixture of experts": two models of the same size trained for different stages of noise. The high noise expert works while the clip is still mostly noise and lays out the composition and the motion. The low noise expert takes over for the second half and finishes the detail. Only one is active at a time, so the pair costs the time of one model but needs both files on disk.
<!-- diagram: a bar of 20 steps; steps 0 to 10 coloured and labelled "high_noise model, sampler 1, adds noise, returns leftover noise"; steps 10 to 20 labelled "low_noise model, sampler 2, adds no noise" -->Where you meet it in a graph
- Two Load Diffusion Model nodes:
wan2.2_..._high_noise_14Bandwan2.2_..._low_noise_14B(t2vori2vversions). - Each model goes through its own ModelSamplingSD3 (
shift, 8 in the 20-step templates) and its own LoRAs. - Two KSampler (Advanced) nodes:
| Field | First (high noise) | Second (low noise) |
|---|---|---|
add_noise | enable | disable |
steps | 20 | 20 (the same) |
cfg | 3.5 | 3.5 |
start_at_step | 0 | 10 |
end_at_step | 10 | 10000 (to the end) |
return_with_leftover_noise | enable | disable |
The first sampler's LATENT goes into the second's latent_image. Both get the same positive and negative conditioning. With speed-up LoRAs: 4 steps, hand-over at 2, cfg 1.
The Wan 2.2 5B model is one model, not a pair, with its own VAE (wan2.2_vae) and one ordinary sampler. In WanVideoWrapper the same split is made with start_step and end_step of two WanVideo Sampler nodes.
Typical mistakes
- The same file in both loaders, or high and low swapped.
- Leftover noise not returned: the second pass has nothing to do and the clip looks unfinished.
- Hand-over steps that do not match: lost or doubled steps.
- Noise added in the second sampler: a noisy clip.
How the app helps
The video lessons build this exact graph and check every field of the table (cfg 1 to 4, hand-over 1 to 15). The lesson names the samplers "first pass" and "second pass".