Not the creator, but from the workflows embedded in the videos, and the creator's other model, it looks like the high noise LoRA is the only one that's needed. High noise is responsible for the large motion and scene composition, so it would make sense that you wouldn't need a low-noise LoRA.

@japarn7 I tried the settings and I still get the issue even without your lora so its something on my end, I think its my computer or maybe the style of anime eyes, I'm using Reze from CSM. How much system RAM and VRAM do you have? I heard you need at least 64GB of ram to have good quality generations. I'm pretty sure my 16gb vram is enough but I do only have 32gb of ram thats always maxing out. I even tried a fresh new workflow and same issue with her eyes with the non GGUF model. I heard face consistency isn't the greatest sometimes so maybe it's just as simple as that and might need a wan lora for the character I'm using, thanks for the response

My deer, I had the same problem. You can try another WAN 2.2 checkpoint. I used smoothMix and eyes were often bad when I tried to fit the whole character in the picture. I changed the checkpoint to DaSiWa WAN (https://civitai.com/models/1981116/dasiwa-wan-22-i2v-14b-lightspeed) made by the mighty deer Darksidewalker. Now the eyes are crystal clear. The bad news: I haven't seen a GGUF model for it. But if you have enough free space on your SSD (about 27 GB), I recommend you try this model in safetensor format. I also have 16 GB of VRAM (RTX 5070Ti) and 32 GB of DDR5 RAM. I can easily generate a 113-frame video (7 seconds) at a resolution of 720x1024 in 12-15 minutes (I can do a little more, but I'm not sure about the time). I use DaSiWa workflows, but I've tried others, and the results were the same. And one more tip: don't use image compression in the workflow, it degrades the quality of the image itself. It's not always very noticeable, but in my honest opinion, the time saved in generation is not worth it. It's better to generate image at a lower resolution than usual, spend a few more minutes on generation, and get a better result (for example, I usually generate images at 1536x1024, but for videos I use 1024x720). Usually, you don't even need to disable the nodes; it's enough to enter the original resolution in these nodes when configuring the generation in the workflow. After all, most workflows have upscalers for the generated video.

@TheCoffeeDeer Oh deer, that model makes the eyes so much better and actually usable. TY. I really liked smoothmix GGUF but this ones perfect for anime. I hope someone makes a GGUF version of that anime model, longer gen times worth it for now

Edit: I found a GGUF verion on hugging face

Did you try to use only High

@hboxgames132 Yes. In fact, the video I posted was generated only with high. A few minutes ago, I tried making a video with and without low noise using the same settings and seed. The only differences I noticed when using LoRA in low noise mode were that the movements became more abrupt and the face quality decreased a bit. Maybe this can be solved by increasing the number of steps. I'll try to make another couple of videos like this, but with a larger number of steps (8 instead of 4). I'll try to write about the results in about an hour.

@hboxgames132 So, more than an hour passed. I would even say about six hours. My ComfyUI updated, broke, I reinstalled it, started generating, it broke again. Repeat several times. All the repairs took several hours. And it still doesn't work completely correctly... Classic.

But I still made this comparison. I posted both versions in this post https://civitai.com/posts/24613174.

In short, there is almost no difference, only the facial movements have changed. In the version with LoRA in High and Low noise, the eyes are more expressive.

@hboxgames132 Yeah using only the high lora with the base model, gguf, or smoothmix gguf checkpoint that I had issues the eyes warping with. It has to do with the checkpoint models and not the lora itself. The high for checkpoints and loras model seem to affect motion, and low affects details. This is kinda confusing but I mixed dasiwa's high checkpoint with the default low wan checkpoint and used them vise versa to test the difference in motion/detail of the eyes. It basically confirmed the blinking problem I was having was a checkpoint issue and I'm now I'm just using dasiwa's high and low checkpoint together since it was specifically made for anime and it works perfectly fine with this lora