Wan2.2_T2V_A14B_VACE-test

[!IMPORTANT] ⚠️ Notice
This project is intended for experimental use only.

This is an addon experiment of Wan2.2 T2V A14B and VACE scopes from Wan2.1 VACE T2V 14B.

The process involved injecting VACE scopes into the target models, using scripts provided by wsbagnsv1.

All GGUF quantized versions were created from the FP16 model using the conversion scripts provided by city96, available at the ComfyUI-GGUF GitHub repository.

Notes

Tested with 2-step High Noise and 2-step Low Noise dual sampling with the LightX2V LoRA, it's working fine in ComfyUI.

There's news where VACE team might have a fix for the color shifting issue to be released (discussions on the Banodoco Discord Server).

Will be waiting for the official fix before testing further.

---

References

🔗 Wan2.2 MoE

• Wan2.2 separates expert models by timestep: The High-Noise expert focuses on generating overall layout and motion. The Low-Noise expert refines textures and details.
• The A14B model includes both High-Noise and Low-Noise experts, which are activated at different denoising stages.

🔗 ComfyUI Wan2.2 14B T2V Workflow Examples

🔗 ComfyUI Wan2.1 VACE Workflow Examples

---