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
Major Update is Here!
Changelog:
Fully retrained on a cleaner, more refined dataset
Fixed issues with catastrophic forgetting
Improved character accuracy and flexibility
Enhanced eye details
Improved overall stability
Note:
For the best results, we highly recommend using rating tags in your prompts. They can significantly improve anatomy and overall generation quality. Since it’s situational depending on what you want to create, start with sensitive if you’re unsure, they work well in most cases.
Updated Character Knowledge:
Honkai: Star Rail → up to v3.7
Genshin Impact → up to v6.0
Zenless Zone Zero → up to v2.2
Wuthering Waves → up to v2.6
FAQ
Comments (16)
I got to know this model in version 6 and thought it was cool, but it still wasn’t exactly what I was looking for. However, version 7 is perfect! Today, it’s definitely my number 1 and favorite model. I love both its style and focus, and the generated images are of great quality. Thanks for keeping the model updated! <3 Today, without a doubt, it’s the best for me.
Thank you so much for liking my model! I’m really glad you enjoy v7. I’ll keep doing my best to keep it updated
out of the models I'm currently using, this is the only one that properly understands the "anime screenshot" concept. I don't like using LORAs for stuff if I don't have to, and this model seems to be much more flexible with styles than other popular models. Almost like RouWei in that respect. Keep up the good work, friend, it does not go unnoticed.
Thank you
Great model as always. Just wonder:
I tried to merge the model with some personal merge model I use that has more mature style (e.g. smaller eyes and slightly longer face),as I feel Raehoshi's style is less "mature", but I notice the as the merge ratio goes up the merged model will forget those newly learned characters, and it is hard to balance the style and knowledge during the merge. Is there anyway to alter the default style of the model toward other model, while keep the knowledge intact?
You can use "mature female" in your prompt to achieve a more mature facial style. If the effect isn’t strong enough, you can increase its weight, for example: "(mature female:1.5)".
If you want to merge models while keeping the original knowledge intact, try using a train difference merge mode or add difference. Make sure to set Raehoshi as the base model (Model A in WebUI/SuperMerger)
@Raelina Thanks for the info! I am using Comfyui tho, will see how to enable the difference merge/add difference mode :)
Will you make a Neta Lumina finetune? (Or NetaYume Lumina finetune) This dataset would be really helpful
Will you add some other games in the future? such as Azurlane, Fate, Arknight
Sure, but I haven't played that game, so I'm not familiar with the new characters. Could you provide a list of characters that are not yet available in this model? or you can DM me if it's too long
@Raelina oh, that pretty hard to list, since there're so many characters these months.
When using v7 or v5.1 for merging, messages like "there are non finite values in key" are printed to the console or the merge process freezes in the following keys:
conditioner.embedders.0.transformer.text_model.encoder.layers.11.mlp.fc1.weight'
conditioner.embedders.0.transformer.text_model.encoder.layers.11.mlp.fc2.weight'
conditioner.embedders.0.transformer.text_model.encoder.layers.11.self_attn.k_proj.weight'
conditioner.embedders.0.transformer.text_model.encoder.layers.11.self_attn.out_proj.weight'
conditioner.embedders.0.transformer.text_model.encoder.layers.11.self_attn.q_proj.weight'
conditioner.embedders.0.transformer.text_model.encoder.layers.11.self_attn.v_proj.weight'
Maybe some part of the text encoder is broken.
v6 is fine.
Can you provide more details about the merge method you used?
When I tested it with weighted sum, add difference, and train difference, everything worked fine without any issues.
It’s possible that the merge model you used has a conflict with v7, or it may not support the merge method you’re using.
@Raelina This is an example of weighted merge between Raehoshi using comfy-mecha.
https://files.catbox.moe/a06caq.png
The image is generated successfully, but the console displays "there are non finite values in key" message similar to the image below.
https://files.catbox.moe/9whkz9.png
Weighted sum, add difference, and train difference will complete the merge although an error message will be output, but merging using comfy-mecha's "Rotate" will be canceled due to an error.
Below is an example of a merge using Rotate.
https://files.catbox.moe/7sheka.png
This merge works fine in V6, Illustrious-XL-v1.1 and NoobAI XL, but switching to V7 produces the following error.
https://files.catbox.moe/xgdzgy.png
ComfyUI workflow images have metadata, so you can reproduce the workflow by dropping them into ComfyUI.
@nuko_masshigura Thank you for your detailed bug report regarding the issue with the model. I tested it using your workflow with the rotate merge method, and yes it does have that issue. I also tested the v7 base (pretrained/untuned version) to trace the source of the error, and the pretrained model didn’t have this problem.
I haven’t tried the rotate merge method before, so I’m not entirely sure what it does. I’ll look into it and try to fix it in the next update. Thanks again for your feedback, I really appreciate it
@nuko_masshigura The new version 7.1 has fixed TE component that caused conflicts when using it for merging