Mask Helper

I had three main objectives when creating this LoRA:

• Get my Min'yō inspired OC wearing a badass mask without AI constantly merging the mask’s horns with her own.
• Generate a proper full gas mask with a hoodie and low twintails that stay correctly outside the hoodie instead of merging with the mask.
• Finally make the AI understand how mouth veils work and stop having the tongue clip through them as if they were part of the chin.

Overall, I was tired of almost every model failing to understand mask physics, the simple fact that masks are objects placed on a character, not living parts of their face with moving eyes, merging with the skin, or causing all the usual glitches.

So what did I do?

I created around a hundred reference images featuring many different types of masks and veils, with special focus on various wearing styles and interactions: properly lifted mouth veils, half-worn surgical masks, masks worn on the side, festival masks on top of the head, and more. I included a wide variety of character archetypes, anthro, girls, androgynous figures, muscular males... to cover as many use cases as possible.

I used AI editing tools, then carefully cropped and selected the best results. For tagging, GPT suggested a structure like this: mask_type:type of mask you want, covers:parts you want to hide, exposes:parts you want to see, tight_fit / loose_fit / attached_straps, and material: cloth / lace / leather / ceramic, etc. (You can try your usual prompt and just use it as a helper)

After testing the LoRA, I noticed significant improvements, which makes me believe the key might hide in better prompting. You can also try some negative prompts like extra eyes, floating mask. I’m still working on it and may come back with even better ideas later. That’s also why I’m sharing this LoRA, to see what solutions the community can come up with and how to improve.

Did it work for me?

Well, it’s starting to become a running gag with my helper LoRAs: I create them to fix a specific recurring issue, only to end up solving other problems instead. I didn’t quite achieve the exact results I was aiming for, but I suspect that’s still more related to my prompting than to the LoRA itself. I look forward to improving it further or even releasing updated versions. Your feedback would be really valuable.

What this LoRA is not: It is not an ultimate all-in-one mask reference. Although I used and created many types of masks, I wouldn’t recommend using this LoRA alone if you’re trying to generate a very specific mask design.

What this LoRA is: It was made to help AI (particularly weaker models) finally understand mask physics, the difference between veils and solid masks, partially worn masks (like masquerade masks or surgical masks), properly generated fully-worn festival masks, masks worn on the side of the head, etc. It also aims to produce cleaner “open mouth” results while wearing harem veils without the usual mess.

Important: You’re free to download, edit, and do whatever you want with it (no commercial use). Credit is always appreciated but not required. If you make merges, please release them under the same terms.