Wan 2.2 - Still Experimental

So, it turns out that Wan 2.2 CANNOT be trained like Wan 2.1. You can train a lora for the high noise model or the low noise model or both, but they have to have different time step ranges. You can also train a lora for the low noise model with the full time step range and run it as a single model. I am currently testing this by training three new versions of this LoRA... a high noise and a low noise model to be used together as well as a low noise model to be used when generating with only the low noise model. More to come!

Triggers: a beautiful girl OR a beautiful young woman OR a beautiful woman... Each will yield different results.

v4.1 - More Diversity!

This version was trained on an expanded, more ethnically diverse dataset. All of the previous improvements still apply.

v3 - Now with nudity!

I noticed that the previous version was not wanting to generate nudity. I refined the dataset and added in some nude images. Problem solved!

Note: Using the word 'woman' vs. 'girl' makes a big difference in the output.

Wan v1 was a little over trained. I revised the dataset and reran the training with new LR and settings to create v2. v2 should be more compatible with combining loras.

This model was trained on 50 images of 50 different women. All pictures are publicly available. No trigger is necessary. Just prompt as usual and turn on the lora.

(If you have an idea or a request, send me a message. I can afford my own coffee.)