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 with multi resolution
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
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 article 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
Updated character knowledge:
Wuthering waves v2.5
Zenless zone zero v2.1
Honkai Star Rail v3.5
Genshin Impact v5.8
Fate: mrs snake
nikke: bready
2 new indie vtuber: nimi nightmare and S***
Note:
This model is still based on Illustrious 1.1 and is a continuation of training from v5.1. The v5 version, which was based on Illustrious 2.0, had stability issues, so for now I will stick with the Illustrious 1.1 base.
Some new characters may be inaccurate, especially in clothing, due to limited fanart available during training.
FAQ
Comments (21)
Gotta wait 2 weeks for 6.0 ......
Honestly, Illustrious or the open source anime model has reach a cap. Illustrious dev team seems have no intention to release newer model or even develop newer model. Although SDXL has so many limitations, it is still the most reasonable model for most personal computer.
I have been trying Novelai V4.5. While it is closed sourced, it does have very strong capacity when rendering details. However there are also issues like tag pollution, backgrounds that lack details, very unstable artist tag combination, and no lora support at all.
I keep dreaming what if we can have a model to combine the advantage of illustrious and Novelai.....But it does not seems like things that can come true. Training new model seems just expensive....
There is Neta Lumina, though. If you can run SDXL comfortably, then you can run that too. It just requires some good finetuning for details and stability, otherwise it has a very good prompt adherence for the model of its size and is better than SDXL architecture-wise and uses Flux's VAE, so the details are better too.
munchkin Thanks for the reply. Yeah I noticed that as well, it just looks like early stage Illustrious 0.1, requiring a lot of finetuning. Compared to IL 0.1 it also has much less visibility. Still a long way to go....
well training a modell that u dont do on ur own pc costs money if u have to rent a powerfull gpu for the training . so sure she wants a bit support. or do u want to spend ur money to give a modell for free away?
Cool model, judging by the preview images, but it costs too much to buy. No money.
3€ in ur view is much ? do u know how much it costs to train a modell to rent a powerfull gpu ? thats more then 3€ or $ ;)
@schneesturmx91988 Not much, but in the country where I am, I can't make such payments. That's why it's very sad. And it's even sadder to endure undeserved rudeness from others.
@GipnoKotov then go on her kofi and not on civitai and if u did it send her a msg with ur transaction code she asks u then if u want the link in email or smth
if $3 is too much for you bro idk maybe fix your life or smth
@akdoism You can't always help the country you were born and live in. Lol
@sansmia i help plenty by my taxes and contributing by working throughout the week lol
I can't seem to gen Piper Wheel without a lora
It seems like v6 is experiencing catastrophic forgetting. This usually happens during finetuning with new concepts, characters, or datasets, where the model sometimes forgets its previous knowledge. Since v6 has already gone through several rounds of finetuning, this issue is more noticeable. I tested v4 and it still retains the Piper Wheel knowledge, so you can try v4 if you want to generate Piper Wheel. Thank you for reporting this, if you find any other characters that are hard to generate, feel free to share your feedback
Does 6.0 include Komano?
It does
Will you make a v-pred version?
Not sure. I tried training a vpred model, but the results still had flaws, so I’ve put it on hold for now until I succeed in training it properly
Due to the 'catastrophic forgetting' issue of v6 you mentioned, will this be fixed in a different version? Which model retains all of the character knowledge before v6?
V4 is the first finetune using Illustrious v1.1 as the base, so it’s essentially the foundation version of this model. Versions before V4 were based on Illustrious v0.1 and NoobAI eps
@Raelina So does that mean v4 retains all the character knowledge or...? I'm not sure if that answered my question.
@madaraxuchiha88 Since you mentioned catastrophic forgetting in v6, I assume you’re referring to my other comment about Piper Wheel and Zenless Zone Zero. Those characters were included in my trained dataset to V4. After V4, I didn’t train them again, and the Illustrious base itself didn’t originally know most Zenless Zone Zero characters. If you’re asking about the original characters that came from Illustrious, I’m not sure, you’d need to test that yourself.