<h2 id="version-2.0-is-out!-i-worked-a-lot-on-the-lora-to-prevent-overfitting-while-keeping-the-appearance-stable.">Version 2.0 is out! I worked a lot on the LoRA to prevent overfitting while keeping the appearance stable.</h2>Read this if you interested: <a target="_blank" rel="ugc" href="https://civarchive.com/articles/29783/lora-training-for-anime-characters">https://civarchive.com/articles/29783/lora-training-for-anime-characters</a>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.<h3 id="some-words-about-the-anima-model:">Some words about the Anima model:</h3>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.<h3 id="pros:">Pros:</h3><ul><li>Incredible tag responsiveness (way better than IL, in my opinion)</li><li>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</li><li>Great output quality even with short prompts</li><li>Natural language prompting worked well</li></ul>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!<h3 id="cons:">Cons:</h3><ul><li>A bit complicated, since you have to include the VAE and text encoder manually</li><li>The models I create tend to drift when using many tags</li><li>No full Hires.fix support yet</li><li>Complicated prompt order (and it seems to matter a lot for output quality)</li><li>More generation parameters to handle (e.g. Shift)</li></ul><h3 id="conclusion-for-now:">Conclusion for now:</h3>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.<hr />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_v170Sampler: Euler aScheduler: <s>Automatic</s> KarrasSamplings 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: 0.3-0.5Extensions (nice to have for little more Details):Adetailer (highly recommended)

Orihime Inoue Hueco Mundo Version - Bleach

Initial Release. Model tends to drift, when using many tags, aesthetics tags help a bit

Files with SHA256: aa0db846124f5cf28c370d0ae9a29ba6a560a535bec03b8b7a6298e4b64347c0