Version 2.0 is out! I worked a lot on the LoRA to prevent overfitting while keeping the appearance stable.

Read this if you interested:

https://civarchive.com/articles/29783/lora-training-for-anime-characters

Important Note: Do NOT use Automatic Scheduler! It will mess up the eye color! Use Karras instead.

Update 14.05.26: Little Note on v2.0 i forgot to mention:
If you use the "hcm-dress" tag, you should add “hcmt-dress” and “torn clothes” to the negative prompt to prevent both outfits from merging.

Some words about the Anima model:

If the image drifts (bright color, oversaturation) use (anime coloring:X.Y) you have to play with the multiplier though...

I spent a lot of time learning about the base model and training the LoRA. Here’s my opinion on Anima.

Pros:

• Incredible tag responsiveness (way better than IL, in my opinion)

• Great outfit awareness during training: with IL, you often have to use tags like “white hakama,” while Anima handled the outfit very well despite the lack of many full-body shots in the dataset

• Great output quality even with short prompts

• Natural language prompting worked well

Before I start with the cons, I want to mention that I’m not entirely sure whether I trained my models correctly. Regardless, I tried many different things, but I couldn’t get better results. Also, don’t forget that Anima is still in beta!

Cons:

• A bit complicated, since you have to include the VAE and text encoder manually

• The models I create tend to drift when using many tags

• No full Hires.fix support yet

• Complicated prompt order (and it seems to matter a lot for output quality)

• More generation parameters to handle (e.g. Shift)

Conclusion for now:

I’ll stick with IL until Anima becomes more stable (or until I find the right training parameters — so I’m not only blaming the base model 😅).

Text for 1.0: I KNOW, I KNOW!

“There are tons of versions of her on Civitai, so why bother making another one — especially when you always say you prefer creating models of characters who don’t already exist on Civitai!?”

But hear me out: I didn’t make another version because the existing ones are bad. It’s just that Orihime Inoue is one of my all-time favorite characters, and I wanted to make my own version of her.

This is only (140 EP) the Hueco Mundo version of her, because that’s my favorite look — though I’ll probably make a Shinigami disguise version at some point too.

Prompt: look at "Trigger Words" on the right side 👉

Changelog: look at "About this Version" on the right side 👉

My Settings (forge)*:

*note, the example images only use high-res fix.

Checkpoint: waiNSFWIllustrious_v170

Sampler: Euler a

Scheduler: Automatic Karras

Samplings Steps: 30 - at least use more than 15! I often see people generate with 8 steps. This it not enough xD (I mean, you should generate you pics the way you want, but the results should be better with more steps)

Hires.fix: R-ERSGAN 4x+Anime6B, 20 Hires steps, Denoising Strength: 0.3-0.5

Extensions (nice to have for little more Details):

Adetailer (highly recommended)