🌸

I created this LoRA because I was getting increasingly frustrated with how anatomical details in this particular area were being generated by Anima checkpoints.

After testing several realistic Anima checkpoints, I noticed that the results were often inconsistent or simply didn't look right to me. So I decided to experiment with training a LoRA specifically to improve this aspect.

And somehow... it actually worked. 💀

What does this LoRA do?

The main purpose of this LoRA is to improve female anatomical detail in realistic Anima checkpoints.

It was specifically created to address an area that I found particularly difficult to generate consistently with the checkpoints I tested.

The goal isn't to completely change the image, but to improve the anatomical structure and overall appearance when that detail is visible.

Dataset

The dataset was generated using Krea 2.

I generated a large number of images while trying to keep the anatomy as aesthetically consistent as possible, including different viewpoints and angles.

The dataset itself was generated with a realistic appearance because my main goal was to improve the results of realistic Anima checkpoints.

Training

I'm going to be completely honest here:

I'm not an advanced LoRA trainer. 😅

I decided to experiment with LoRA training directly through CivitAI because I wanted to see if I could solve this particular problem myself.

The resulting LoRA worked well enough for my own workflow that I decided to publish it.

So if you're experienced with LoRA training and have suggestions for improving the dataset, captions, training settings, or anything else, I'd genuinely love to hear them.

Compatibility & Testing

This LoRA was primarily tested with realistic Anima checkpoints, and this is where I recommend using it.

It appears to be compatible with realistic checkpoints from the Anima family.

I also tested it with anime-style checkpoints.

It can work, but there is a downside: the result may shift the image toward a 2.5D / 3D appearance, since the LoRA itself was trained using realistic imagery.

Because of that, I do not recommend using it with strongly anime-styled checkpoints.

If you're using it together with an anime workflow, I recommend applying it through ADetailer rather than affecting the entire image.

In fact, for my own workflow, I consider using it with ADetailer practically essential if you want the best results.

Recommended Settings

I recommend starting somewhere between:

0.6 – 1.0

Lower weights may have very little visible effect depending on the checkpoint and image.

For the best results, I strongly recommend using the LoRA through ADetailer, especially when the affected area is relatively small in the image.

As always, the exact result will depend heavily on the checkpoint, prompt, composition, and other LoRAs being used.

Limitations / Known Issues

There are a few things I've noticed during testing:

• Primarily intended for realistic Anima checkpoints

• Not recommended for strongly anime-styled checkpoints

• May slightly affect lighting

• May slightly change skin texture

• Results can vary depending on the checkpoint and composition

• The belly button can occasionally become misaligned in certain poses

The last issue is caused by one particular image in my current dataset. I already know which image is responsible, but I haven't corrected the dataset and retrained the LoRA yet.

If you encounter the misalignment, simply changing the seed can often produce a better result.

I'll probably fix that dataset image and retrain the LoRA at some point, but for now I wanted to publish the version that already works well for my own workflow.

Trigger Word

The trigger is simply:

pussy

You don't need anything complicated. Just include it in your prompt when you want the LoRA to activate.

Final thoughts

This was honestly more of an experiment than a carefully planned professional LoRA project.

I originally made it for a very specific reason: I like creating cosplay images with my own model, and my workflow relies heavily on Anima and Krea 2. I wanted to improve an aspect that was consistently frustrating me across the checkpoints I was using.

So there's a good chance this LoRA may not behave exactly the same way in your workflow as it does in mine.

For me, however, it solved a problem that was bothering me enough that I'll probably be using it in all of my future generations where it's relevant.

One important recommendation: use it with ADetailer. In my testing, this is practically mandatory if you want consistently good results. Applying it directly to the entire image can introduce unnecessary changes to things like skin texture, lighting, or the overall style.

There is also the known belly-button alignment issue in certain poses, caused by one image in the dataset. Changing the seed usually helps, and I'll correct the dataset when I get around to updating the LoRA.

This isn't meant to be a perfect, universal LoRA. It's basically the result of me trying to fix a problem in my own workflow and being surprised that the experiment actually worked. 😂

Maybe it'll work perfectly for you.

Maybe it'll need some tweaking.

Maybe it'll completely ruin your generation. 😂

Either way, if you test it, I'd really appreciate feedback — especially from people who know more about LoRA training than I do.

Suggestions for improving the dataset or training are very welcome! 💚