Raehoshi Illust XL
an enhanced iteration built upon the Illustrious XL model. It aims to elevate the visual style by addressing some of the limitations in the original, such as oversaturation and artifact noise. While these issues are not entirely eliminated, noticeable improvements have been made. The goal is to deliver a more polished, balanced output while staying true to the strengths of the base model.
Why Early Access?
Early access helps keep the project going. I don’t have my own GPU, so all training is done through rented cloud GPUs and that gets pretty expensive. By getting early access, you’re directly supporting the development of my models and helping me keep improving them. If you'd like to support me further, you can also buy me a coffee on Ko-fi! Every bit of help means a lot and keeps the future updates coming.
Recommended setting
Positive prompt :
masterpiece, best quality, very aesthetic, absurdresNegative prompt :
bad quality, worst quality, jpeg artifacts, sketch, bad anatomy, signature, watermarkSteps : 25+
CFG : 5-7
Sampler : euler a or dpm++2m karras (euler for vpred)
Standard resolution :
832 x 1216, 1216 x 832, 1152 x 896, 896 x 1152, 1344 x 768, 768 x 1344, 1024 x 1024High resolution :
1024 x 1536, 896 x 1536, 1536 x 1024, 1536 x 896Hires.fix Setting:
Upscaler : 4x Foolhardy Remacri
Hires step : 10-15
Denoise : 0.1-0.3
Special Tags
Quality Tags:
masterpiece
best quality
good quality
average quality
bad quality
worst quality
Rating Tags:
safe
sensitive
nsfw
nsfw, explicit
Aesthetic Tags:
very aesthetic
aesthetic
displeasing
very displeasing
Training Details
The model was developed using a two-stage fine-tuning process. In Stage 1, new series and characters were introduced into the model. Stage 2 focused on fixing issues and enhancing the overall style for improved output.
Stage 1
Dataset : v1-31k, v2-37k, v3-34k, v4-60k, v5_v5.1-18k, v6-15k, v7-39k, v8-41k, v9-30k, v10-30k, v11-29k
Hardware : 2x A100 80gb, v3, v4, v5, v5.1-2x H100 80gb, v7,v8, v9, v10-RTX PRO 6000
Batch size : 32
Gradient accumulation steps : 2
Learning rate : 6e-6
Text encoder : 3e-6
Epoch : 15
Stage 2
Dataset : v1-2.5k, v2 and v3-2.3k, v4-2.5k, v5-2k, v5.1-1.8k, v6-1.5k, v7-1.7k, v7.1,v8-4.1k, v9-1.9k, v10-2.4k, v11-3k
Hardware : 1x A100 80gb, v7_v7.1,v8, v9, v10-RTX PRO 6000
Batch size : 48
Gradient accumulation steps : 1
Learning rate : 3e-6, v5.1-2.5e-6
Text encoder : disable
Epoch : 15
List of New Series/Characters Trained:
Zenless Zone Zero
Wuthering Waves
Honkai: Star Rail
Genshin Impact
Arknights: Endfield
Umamusume
Azur Lane
Arknights
Fate/GO
Dandadan
Make heroine ga oo sugiru
Kusuriya no Hotorigoto
Hololive from justice and dev is
Indie Vtuber Dooby, Yuuki Sakuna, Nimi Nightmare, and S***
100 girlfriends who really love you
Haite kudasai takamine-san
Alina clover
Nikke: bready and little mermaid
Kpop Demon Hunters
Full character list are available here
For character trait details prompts, please refer to the Danbooru site for accurate tags and references.
License
Special thanks to Joe for supporting my works
Special thanks to Juno for supporting my works and help me with early tester
Description
New Training Methodology
This update introduces a refined multi-resolution training pipeline designed to enhance output quality:
Multi-Resolution Scaling: Optimized across low, medium, and high-density datasets.
Enhanced Performance: Significant improvements to structural stability and fine-grain detail.
Refined Fidelity: Superior character accuracy achieved through adaptive resolution training.
Expanded Knowledge Base (Up to April 2026)
The model’s character and lore library has been updated to include the latest data for:
Arknights: Endfield
Honkai: Star Rail
Zenless Zone Zero
Genshin Impact
Wuthering Waves
The full character list can be accessed here:
FAQ
Comments (19)
Any plans on an Anima version?
I haven't decided yet. Improving character knowledge through model training is quite expensive, and the current licensing restrictions make me hesitant to commit. I’d definitely be open to it If there’s interest in sponsoring the training costs, I’d be happy to try it
@Raelina If you can share, how much would it approximately cost?
@GSLinux It depends on the size of the dataset. For example, training with 10k images would cost approximately $100 to $200, or perhaps more. This estimate covers the entire development cycle, not just the training itself including scraping, captioning, cleaning the dataset, performing POC tests, etc. Please note that this is for a full model training, which is much more intensive than training a lora.
@Raelina I sent a separate message. Thanks!
@Raelina That sounds extremely dubious.
Civitai has a model creator named Crody who creates excellent models and releases them without early access. He has significantly more models to release. Users contribute to the costs themselves, and the donations generate significantly more revenue and are significantly faster.
So your statement sounds extremely dubious. I think you need to reconsider that approach.
@Gipno Could you clarify which part of my statement seems dubious? You cannot compare my models with Crody, as we use entirely different methods. Crody primarily focuses on checkpoint merging, which doesn't require renting expensive high-end GPUs for days.
My models are created through full native training, which is an extremely resource intensive and costly process. Every creator has their own way to create a model, and early access is what allows me to keep doing full training. Comparing a merge to a full native train is simply not an apples-to-apples comparison
@Gipno whether to use early access or not is a creator's choice, and this is the approach that works for my development process
I found v10 quite interesting; I'm still testing it, but it's pretty good.
v10 is probably the best Epsilon model so far for me. I started using your model with the vpred 1.0 version and have kept using the 2.0 update too. I honestly hope you keep updating the vpred version too and keep the same-ish style as it's the only model I've had good success in getting what I want out of it. So far, v10 EPS is great though, thank you for the models and the fact that you keep introducing new characters each time.
The author's model is really impressive, new characters can also be created directly without using the Lora model. I hope it can continue to be updated, come on, author
So, which artist styles are supported?
The model is primarily focused on character knowledge, not specific artist styles. Although some artist tags exist in the dataset, they aren't fully trained enough to guarantee a consistent style. That's why I haven't provided an artist list
I love how clean the generated result are. even with quite a badly trained loras.
V10 is clear cut from v9.1 and 9.0, and arguably better than vpred 2.0 if we just count output quality generated.
I'm really glad to hear that you're enjoying the results. Thank you for the detailed review and for comparing the versions. Happy generating!
Thank you for this. I realize a terabyte of loras just wasn't worth it LMAO
使用v10生成的图像光照在人物身上似乎有些过亮，我想要生成黑暗的氛围有些困难，不知是否只有我一个人有这个问题...
Another great version, the new characters work great, I haven't done much testing, because my time has been limited for the last few months. Either way, for what I tested, it seems certain details for some characters still got lost. Ex: Alice (Genshin) her hat doesn't work anymore, Kpop Demon Hunter girls' hairs are hit or miss, even with minimal or a lot of prompts. Like Rumi braid looks more like a high ponytail, something that didn't happen in 9.1.
I have a theory or suggestion, what if in the next training you add a small number of images of all the characters you have trained so far? Maybe that will help the training to not fully overwrite the old character data you trained, maybe 20 or 50 images per character is enough? I'm not sure, but just wanted to share my theory. Unless you already do that, then ignore me lol
Either way, the data is still there and it's working for most characters, my testing hasn't been that intensive. But overall, it will work for most people, so thanks for the great model! 👍
总感觉对画师的还原度不够好，受模型默认风格影响很大