File size: 2,111 Bytes
2e08fdf
 
 
 
 
 
 
 
3d9f50f
2e08fdf
7d15992
2e08fdf
7d15992
3d9f50f
 
 
7d15992
2e08fdf
c5d3f92
2e08fdf
3d9f50f
2e08fdf
 
 
114ec3f
2e08fdf
a1e7411
2e08fdf
 
a1e7411
 
2e08fdf
 
edeba03
 
 
 
8b886c5
2e08fdf
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
---
license: apache-2.0
---

> [!IMPORTANT]
> โš ๏ธ **Notice**  
> This project is intended for **experimental use only**.

This is an addon experiment of **[Wan2.2 T2V A14B](https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B)** and VACE scopes from **[Wan2.1 VACE T2V 14B](https://huggingface.co/Wan-AI/Wan2.1-VACE-14B)**. 

The process involved injecting [VACE scopes](https://huggingface.co/Wan-AI/Wan2.1-VACE-14B/blob/main/diffusion_pytorch_model-00007-of-00007.safetensors) into the target models, using scripts provided by [wsbagnsv1](https://huggingface.co/wsbagnsv1). 

All [GGUF quantized](https://huggingface.co/docs/hub/en/gguf#quantization-types) versions were created from the FP16 model using the conversion scripts provided by city96, available at the [ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF/tree/main/tools) GitHub repository. 

## Notes

Tested with **2-step High Noise** and **2-step Low Noise** dual sampling with the [LightX2V LoRA](https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Lightx2v), it's working fine in [ComfyUI](https://docs.comfy.org/). 

There's news where VACE team might have a **fix for the [color shifting](https://github.com/ali-vilab/VACE/issues/44) issue** to be released (discussions on the [Banodoco Discord Server](https://discord.com/channels/1076117621407223829/1342763350815277067/1399409879017459903)).  

Will be waiting for the official fix before testing further. 

---

## References

๐Ÿ”— [Wan2.2 MoE](https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B#1-mixture-of-experts-moe-architecture)

>- **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.

๐Ÿ”— [ComfyUI Wan2.2 14B T2V Workflow Examples](https://docs.comfy.org/tutorials/video/wan/wan2_2#wan2-2-14b-t2v-text-to-video-workflow-example)

๐Ÿ”— [ComfyUI Wan2.1 VACE Workflow Examples](https://docs.comfy.org/tutorials/video/wan/vace)


---