Awesome. I think many lora models have introduced same face syndrome because they use AI generated images from Illustrious/SDXL/PNY/NoobAI and so on instead of using actual people from like stock images and such. WAN22 by itself has varied faces but the moment you introduce some loras that facial variety collapses and low and behold the ol' same face from SDXL reappears or one that looks VERY close.

One work around I've discovered is to use Res5s with linear quadratic and suddenly I'm getting different faces. Even in illustrious and noobAI. SDXL and Pony I don't use anymore since nothing seems to solve the same face syndrome there. Even loras don't work half the time. So I can't comment on how well it'll work with those models. But linear quadratic (It can be difficult to use) with a Res5s (3s works too, 2s has less impact but 3s and above work well) sampler seems to actually make faces that have unique features.

It takes some trial and error to get linear quadratic to function without it falling a part but if you can get it working you'll see that prompt adherence is better too.

It has to be used with a positive weight, right? The SDXL 'sameface' LoRA i've been using needs to be set to neg

Hi, thanks for making this Lora, could give an example of how you'd use it in the general description section? Thanks again

Thanks alot! with that large training set this sure should fix it

Could someone kindly explain to me better what this Lora is supposed to do, because I'm struggling to understand it?

If you didn't clearly caption each different person with some unique trigger, I doubt it will make more different faces, it will most likely create a blend of all those face and create another unique one, didn't tested yet, but seeing your demo images already confirm my bias

For the men, were any of them younger and pretty? There seems to be a bias on this site toward old or homely men.