Wan2.2模型配置器

Wan2.2 is the model that split itself in two. Its architecture runs two separate transformers - a high-noise model that handles motion and composition, and a low-noise model that refines detail - and that split lands right in the middle of this node. Unlike every other model picker in the pack, Wan22ModelNode makes you confront it: min_t and max_t are required inputs that define which noise band you're training on.

This isn't pedantry; the community spent real time getting confused by the two-pass design. The KB's Wan video notes call the low/high-noise split "a source of significant community confusion," with people speed-LoRAing one pass and not the other. The same split applies to training.

How it works

You give it a checkpoint directory (ckpt_path, which needs at least the VAE and config files), and the node builds a {"type": "wan", ...} config with a min_t/max_t noise range baked in. The tooltips spell out the two presets:

• Low-noise model: min_t = 0, max_t = 0.875
• High-noise model: min_t = 0.875, max_t = 1

The range controls what portion of the noise schedule your training samples. Set it wrong - say the full 0–1 on a low-noise model - and you're training on a band of timesteps that model doesn't own. The transformer_path override also follows the split: for ComfyUI-format weights you point it at the specific subfolder (.../low_noise_model) or the matching safetensors, plus a separate llm_path for the UMT5 text encoder.

The inputs that matter

• ckpt_path - required checkpoint directory (VAE + configs at minimum).
• min_t / max_t - required, the noise band. Low-noise: 0 → 0.875. High-noise: 0.875 → 1.
• transformer_path - optional override; can point at a low_noise_model subfolder or a ComfyUI-format safetensors.
• llm_path - optional text-encoder path for ComfyUI-format checkpoints.

Output: model_path → ModelConfig.

Install

Pack-wide install, as with every node here - ComfyUI Manager (search "Diffusion_pipe_in_ComfyUI") or:

cd ComfyUI/custom_nodes
git clone --recurse-submodules https://github.com/TianDongL/Diffusion_pipe_in_ComfyUI.git
cd Diffusion_pipe_in_ComfyUI
git submodule init && git submodule update
pip install -r requirements.txt

Linux/WSL2 only. Heavy deepspeed/diffusers deps included.

Where people get burned

Mostly on the two-pass reality. If you train only the low-noise model, your LoRA refines detail but inherits whatever the high-noise pass does - which is fine if you like the base motion, but a limitation if you wanted full control. Community wisdom on 2.2 training is to train low and high (reported ~90 minutes total on a 5090), but the simpler on-ramp is training 2.1 and using the LoRA on 2.2, since 2.1 LoRAs are forward-compatible. Also double-check the band against the checkpoint you actually loaded: training a high-noise model with low-noise min/max values produces garbage that's hard to diagnose, because the run looks perfectly healthy in the logs.