ATTENTION! This is my first Wan 2.2 LoRA and is mainly for testing and sharing ideas about training a Wan2.2 LoRA, I am still testing different training configs and combination, I don't believe this is the best result Wan2.2 can achieved. Please let me know if you have any experience on training Wan2.2 LoRA.

ComicBook-Wan2.2-T2V-A14B. You can use MediaInfo to check the generation info in the video. The workflow I used is the official one but adding one LoRA after loading the ckpt.

wan22-style7-comic-sel-2-000020.safetensors is for low noise

wan22-style7-comic-sel-highnoise-1.safetensors is for high noise

As we all known,Wan 2.2 T2V-A14 contains two parts ,the high noise and the low noise, both two ckpts share the same structure as Wan 2.1 T2V-14. For my personal generation experience , the high noise is mainly focus on the general of the video like the motion of charactor and camera, while the low noise is mainly focus on the details like texture. Because Wan 2.2 is a MOE model, it is difficult to finetune two parts together directly by using an consumer GPU.That's why I try to make LoRA separately. As for the training tools, because the structure has not changed, temporarily you can train it by diffusion-pipe/musubi-tuner etc.

What I have tested:

1. Low Noise LoRA trained and apply only on Low Noise

For my experience, sometimes it get a great result like the gif below, but sometimes the art style is not that correct.

2. High Noise LoRA trained and apply only on High Noise

Because this LoRA is only trained in image, if we only use this LoRA with a high strength, It will decrease the movement of the video(the same as Wan2.1). I haven't test the only video dataset, I will train a new one just in video tonight.

3. Combination High Noise and Low Noise Together(This one)

Not very stable, sometimes it will not have the result in right style. But in most cases, it can generate a great result.