Happy Halloween!
V1 ANIMA
The first version is for ANIMA. It works stably. Through prompting, the model perfectly creates the effect of dress transparency in intimate areas. For this, use the following tags:
Covered nipples, and/or
Covered pussy,V0.8 IL
Here is a sheer ghost dress that I prepared for Halloween with a slight delay. Unfortunately, not everything works perfectly. I tried to achieve a more transparent effect, but unfortunately I was unable to implement all the features of the dress in one version. However, with some adjustments, you can get the desired result. I plan to release a new version later, which should provide more transparency for covered nipples
Usage tips
DPM++ 2M Karras 30 steps = most stable settings
Model weight: 0.5 - 1.0
Hires denoising strength: 0.2-0.3
For NSFW tags, reduce the model weight to 0.5-0.6, and then use hires.fix if necessary.
Or, when using hires.fix, reduce the model weight to 0.5-0.6. (especially for adding uncharacteristic details). More details below.
Solution for NSFW (covered nipples or covered pussy):
The main problem is that by default, the dress generates pasties and ignores the "covered nipples" and "covered pussy" tags.
Tests have shown that the problem can be solved by reducing the weight of LoRA. So you should add the following tags to the prompt:
For covered nipples: covered nipples, nipples,
For covered pussy: covered pussy, pussy, cameltoe
You need reduce the model's weight to 0.5-0.6 during generation. It turns out that in this case, the nipples generated are quite satisfactory. If necessary, use hires and set denoising strength to 0.2-0.3
Another way to fix this for version 0.8 is to put the "covered nipples" and "nipples" tags together in the prompt and use hires.fix immediately after generation. Denoising strength for hires.fix should be set between 0.2 and 0.3. Before using hires.fix, it is recommended to reduce the model weight to 0.5-0.6.
Similarly, for the tag "covered pussy," you need to add the tags "covered pussy, pussy" to the prompt (+"female pubic hair" if desired).
The third way is to use inpainting right away where you want it.
(This advice is most applicable for local generation via A1111 or Forge. It may be suitable for online generation if you can generate adult content. I can't check online generation on Civitai)
It is worth noting that this method, and sometimes even normal generation, slightly changes the ghost print. The idea is to have two eyes and a blush stickers on the print. But sometimes a print that looks like a mouth or nose appears. Overall, this is not critical, and you can decide for yourself whether to leave it as is or not. You can fix this by regenerating the art with the correct settings, inpainting, or manual refinement.
But overall, the model is quite stable. If you don't want to add nipples or pussy under dress, then all of the above steps are unnecessary.
Description
FAQ
Comments (8)
This is a weird problem. For some reason, civitai is showing this model's size as 43.8KB, when it's actually 43.8MB. I've never seen any other files on this site just display the wrong unit like that
Hmm, that's strange. I train LoRAs for ANIMA locally. On some models the size displays correctly, on others it doesn't. The files are fine on my computer. I reuploaded the same files and now the size is correct, so it's definitely a site issue
I suspect the connection to the server was unstable. The system might analyze the file size before the upload is complete, treat the partial size as correct, and not update it later. Reuploading solves it. I'll watch for this next time. Thanks for the heads up
Nice, and Halloween is just around the corner :P
Any reason the Illustrious version is v0.8? Are you planning to update it or is decent enough as is?
I usually mark a model as v0.8 when I train it on a limited dataset. In this case, the model works quite decently, but in some cases it may need additional tweaking. In the future, I may update such versions to 1.0 with an improved dataset. Based on many tests, the current version works pretty well with the right prompt, and I wrote about the model's shortcomings in the description. I might try to update the model to version 1.0 by Halloween. Though I'll probably also want to make something else Halloween-themed.
@Shirowalker Ah, that makes sense. Credit to you for actually doing that rather than labelling it as v1 and calling it a day. I'll definitely have a play around with it soon then and see how it handles.
Is there anything i can do image-wise to help bolster your dataset? I'm always experimenting with tags to reinforce a lora's concept anyway, so if there's anything i can try to look into for you while i'm generating i'd be more than happy to. (⌒ω⌒)
@NeilSylver Thank you for showing interest in my model. I would really appreciate it if you shared your results on the model page and left a like :)
@Shirowalker I definitely will, don't worry. I generally pick out at least 10 images from all my gens and post them along with my feedback and thoughts for every lora I use. I also picked up a few of your other outfits too so I might pop up on those pages as well :P
(I saw all your OC's too, i'm sitting on a few from other creators already but now I know where to come when i need more) Thanks for the reply, and the lora! I'll be back soon! (⌒▽⌒)☆








