EnhanceMotion

Improves animation speed. For FastMix use.

SelectModel

Switches between the EasyWan22 standard FastMix model and base models. Base models do not have CFG fixed at 1, so you need to select something other than Nag in the adjacent TextEncoder.

TextEncoder

• Kijai: Can be used stably.
• Nag: Allows negative prompts even with Cfg 1 high-speed generation.
  • However, this produces different results from other TextEncoders, for better or worse.
• Native: More precise than Kijai with fp8_scaled, but VRAM-related issues are common.
  • Use with ample VRAM headroom.

TorchCompile

• Disable when older GPUs like Geforce RTX 20x0 or earlier cannot properly use compilation optimization.
  • In such cases, you may also need to set attention_mode to spda in ModelLoader.

ModelLoader

• Lora(1~5)-(High|Low), Lora(1~5)-(High|Low)Strength

  • Set the LoRA to load manually and its strength.
  • Enter so-called trigger words in PositiveInput.

• BlocksToSwap

  • If you don't understand VRAM usage, leave it at 40.
  • Reducing this value allows faster video generation.
    • Choose whether to use excess VRAM for high-precision models, increased video resolution/duration, or faster generation.
  • offload_img_emb, offload_txt_emb, prefetch_blocks are also swap-related settings.
    • Some environments may become faster with prefetch_blocks set to 1. It was slower on my setup.

• base_precision

  • Changing from fp16_fast to fp16 slows down speed but improves accuracy.

• quantization

  • When using fp8 models instead of GGUF, match the model's quantization.

• load_device

  • If you have ample VRAM, loading to main_device might provide speed improvements.

• attention_mode

  • For older GPUs like Geforce RTX 20x0 or earlier that cannot use SageAttention, change from sageattn to sdpa.

• FastMix-(High|Low)

  • Download and use FastMix models with different precisions using Download/diffusion_models/FastMix/Wan22-I2V-FastMix_*.bat.

• Base-(High|Low)

  • Download and use Base models with different precisions using Download/diffusion_models/Base/Wan2.2-I2V-A14B-*.bat.

FAQ

I want to use base models instead of FastMix

Select Base in SelectModel under Model, and choose Kijai in the adjacent TextEncoder. More detailed settings are available in the Base-Sampler subgraph at the bottom left.

I haven't fully tuned the parameters locally and haven't improved compatibility with EndImage. If you have recommended settings that work well with any combination of features, regardless of LoRA or EndImage presence, please let me know.