Zniya is an original fictional character LoRA designed for versatile character generation across photorealistic and 3D/CGI styles
Experimental Character LoRA
This is an experimental character LoRA designed to maintain a consistent character identity across different visual styles, ranging from photorealistic images to stylized 3D/CGI character renders.
One of the main features of this LoRA is its full-body tattoo design, extending from the neck across the torso, arms, legs, and feet. The tattoo appearance is trigger-controlled, allowing the character to be generated without tattoos when the trigger word is omitted, providing greater flexibility for different types of generations.
This release is shared as an experimental/community test version. While the model has been tested across different prompts, poses, outfits, environments, and visual styles, some inconsistencies or artifacts may still occur depending on the generation settings and workflow.
I’m sharing it with the community so others can experiment with it, explore its capabilities, and help identify areas that could be improved in future versions.
Feedback, generation examples, and suggestions are highly appreciated. If you find interesting results, unexpected behavior, or particular settings that work well, feel free to share them.
Full-Body Tattoo Trigger : Zn1ya, The most important element is her body, which is completely covered in intricate, colorful tattoos in a mix of black, red, green, and orange. extending from her neck down to her feet, Her arms, legs, and torso,
Recommended Settings:
Strength:
0.8 – 1.0Sampler:
er_sdeScheduler:
beta
⚠️ Important Note
The dark/black shading around the eyes is a permanent part of the character design and is intentionally included in the LoRA. It cannot be reliably removed through prompting. If you generate Zniya without the tattoo trigger, the full-body tattoos will not appear, but the dark eye-area shading will remain as part of her character appearance.
The tattoo patterns are not fixed. The tattoo design may vary from generation to generation, with different patterns, details, colors, and arrangements appearing each time. This variation is intentional and allows for more diverse results.
Description
Comments (6)
Cool concept character - thanks for sharing!
Just posted a "quick" image set for comparison (okay okay... it took one hour and most of that just to make her lift her hair for the back view...) so it's better to see the consistency of the tattoo. Especially the two 3/4 rear views where she just moves her legs, the pattern on the butt visibly changes, so it's evident that the engine doesn't understand the concept 100%.
But: it looks pretty solid overall. So... good job :)
On the topic, i think it's also relatively easy to make this character believable, because the full-body tattoo is basically just a complete random layer of forms and colors. And TBH i could get a very very similar tattooed woman just by removing the LORA and keeping the character description with its trigger word chain.
HOWEVER ---> Have you ever thought about trying something like these:
https://civitai.red/posts/31387228
https://civitai.red/posts/31400643
https://civitai.red/posts/31409111
https://civitai.red/posts/31487249
Now these are all manually prompted so you can't actually move the character without everything being interpreted and projected differently; but i wonder how a clearly structured, actually realistic ink would work. Perhaps getting reference images would be the hardest part.
Thanks! :) And yes, you noticed exactly one of the things I’m still working on.
In my own tests, I also noticed that the model sometimes struggles to keep the tattoo pattern fixed in the exact same place, especially when the pose or camera angle changes. I actually mentioned this limitation in the description/notes as well.
This character is a bit more challenging than a typical character because the concept is that her entire body is covered in tattoos. That’s actually the main reason I wanted to work on this concept in the first place ... I wanted to see how well the model could learn and reproduce a full-body tattoo design consistently from different angles.
I’m going to keep testing and training it further. I’m planning to push the training up to around 10k steps and see if I can improve the tattoo consistency and get it as close to 100% consistent as possible.
I know that getting the exact same tattoo pattern across completely different poses and viewing angles is quite challenging, but I want to see how far I can push it. :D
And honestly, your comparison set is really useful for testing this. The two 3/4 rear views are a great example of where the model still needs improvement.
Thanks again for taking the time to test it and share the results! 🙌
@LatentCraft If you tab between the second 3/4 rear view and the full rear view, keep a close eye on the left glute. You'll notice that one piece of the tattoo seems to stay in almost the exact shape and place, while everything else changes.
Now this is apparently due to how Krea2 does these skin details. You'll see the same on elements like freckles or pubic hair - if the character pose shifts a few millimeters, but the detail is 100% identical, suddenly looking like an overlay which didn't move. The specific freckles or pubes stay where they were in the previous image (you only see a different section of it, applied to the shifted area of the body).
Basically these general details are just projected from whatever template fits in the datasets. More specific details might work differently.
What you can try, is to not caption the tattoos at all in training. Then it should treat it the same like skin, and render (theoretically) all the same in any condition. Thus you also wouldn't need trigger words.
@LurkingAI Yeah, I actually understand what you mean, and that’s a really interesting idea.
The random/full-body tattoo approach was intentional because I wanted to see if the model could learn a character whose entire body is covered in tattoos and still keep the overall design consistent across different angles and poses.
But I agree that a clearly structured and realistic tattoo design would probably be a much better test of whether the LoRA can actually learn the tattoo as a specific concept rather than just learning it as a layer of colors and shapes.
The reference images would definitely be the difficult part. Ideally, I would need the same tattoo photographed from multiple angles and with different poses so there is enough information for the model to understand how the ink wraps around the body.
I might actually try this as a second experiment after I finish pushing this one further. It would be interesting to compare the two approaches and see how much difference the tattoo structure makes.
And yes, I’ll definitely keep your examples in mind. Thanks for the idea! 🙂
@LurkingAI Ah, now I see what you mean. That actually makes a lot of sense, especially after looking closely at the left glute between those two views.
I hadn’t considered treating the tattoo as a skin detail rather than explicitly captioning it as a concept. That’s actually a very interesting experiment.
For this LoRA I did caption the tattoos, mainly because I wanted the model to learn the tattoo design specifically and keep it as part of the character. But your idea makes me curious about what would happen if I completely leave the tattoos out of the captions and let Krea2 treat them more like an inherent skin characteristic.
I’m definitely going to test that approach. It would be really interesting to compare a captioned-tattoo version against a non-captioned version and see which one gives better consistency across different poses and angles.
If the tattoo behaves more like skin and stays consistent without needing a trigger word, that could actually be a better approach for this particular character.
Thanks for pointing this out — I think this is worth experimenting with. 🙂