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
Update Trained Character up to Jan 2026:
Arknights: Endfield
Umamusume
Zenless Zone Zero
Wuthering Waves
Honkai: Star Rail
Genshin Impact
Kpop Demon Hunters
For complete character list go here:
FAQ
Comments (23)
Thanks for the new version. Are there any conflicts with Laura? Do they support lora nai and pony?
Loras from NoobAI should work properly, but I’m not sure about Pony
I been wondering, you train on top of your previous trained model? or your just train on top of the base again but with the new data every time?
Either way, awesome work, your training always produce good results!
Thank you so much! I use a continued training approach. Each new version is trained on top of the previous one (e.g., v8 is trained from v7, v9 from v8, and so on). Since I use a full/native training method, this allows the model to keep evolving and improving. I'm glad you're enjoying the results
@Raelina Interesting, but you retrain old data + new data on top or just add the new data by itself to the previous trained model? Because by the time you train again, new characters artwork its added to Danbooru, but I assume it's easier for you to select just a certain amount of artwork per character and then just train that over.
Anyways, sorry for the questioning, I was just curious since the model has kept a consistent quality while adding more characters.
Anyways, keep up the awesome work! 😁
@darionk No worries. I'm happy to share the process. I mainly train on just the new data to keep the training efficient. Since I have a limited budget, training the entire Danbooru library every time isn't feasible, so I carefully select specific artworks per character.
I only retrain old data if there are significant metadata changes. For example, 'sunna (zenless zone zero)' was originally trained as 'chinatsu remiel' in v8, once the name was officially changed, I retrained that specific data for v9 to keep everything accurate. Thanks for the curiosity and the support
@Raelina Yeah, I was wondering if you would keep Sunna old name or not, but your image preview answered my curiosity when I checked it. And your method makes sense for efficiency, and I thought as much, but wanted to ask either way, since you never know if someone does something different XD
But thank you for indulging in my curiosity and explain the details of your training!
Will you keep going on v-pred version?
Not sure yet, but I’ll try to continue maintaining the v-pred version. Stay tuned
is pretty bad at doing feet and hands. other than that its good
Thank you for the feedback
Can you upload this checkpoint to https://tensor.art/?
Please,their face fix/hand fix ADetailer are great.Far more better than CIVITAI.
@Raelina Thx
yeah it is good but not NSFW support, I mean cleavage is common thing
First of all, I want to again appreciate your work and thank you for what you are doing, and I wanted to do some feedback:
I been playing around with the model and I have noticed some stuff, although I dunno if its due to not proper prompting or something else, but here are my observations:
1. Older characters (v8 characters for example) are harder to bring out correctly with the same amount of prompts as in v8. You either need to add more or get a lucky seed.
2. Some of the new characters, like nangong_yu, are hard to get properly, her hair in particular. And noticed some stuff like that for other characters, it feels less consistent than previous versions. But maybe Nangong has less images.
So I wonder if after several retraining over, if older trained data has been affected. That brings me to some questions regarding that:
1. Have you tried to train from zero? I assume you avoid it because it will take more time and unfortunately more money, but that brings me to the next question.
2. Are the characters database different amounts or you try to keep them inside a certain amount? I know some characters have way more art than others, but maybe capping the data to 100 or less images per character might help? That way all data stays in a certain size, but dunno how useful that could be.
3. Have you tried training only certain Unets? I ask because it can shorten the training time and I have noticed with Lora and Mergers, that some Unet do have most of the character data, the rest have other less important info for characters. Maybe limiting the training to IN08, MID, OUT00, OUT02 would be enough? BTW I took out OUT01 because that is the one that carries a lot of artstyle info, so training OUT01 does change the style of the model but I know it has some color info, so maybe training only attn in OUT01 is enough?
Anyways, sorry for the long comment, maybe a lot of my issues with v9 are just me not prompting correctly and the model works well. This is just based on what I have noticed and wanted to share my observations. Again, thank you for your hard work! 😁
Thank you for the detailed feedback! This is very helpful for me.
Regarding your observations:
1. You're right, v9 is experiencing some catastrophic forgetting where older characters are becoming harder to generate. I’m currently working on a fix for this and hope to have it ready within the week.
2. Character like Nangong Yu had fewer images during training, which is likely why she’s harder to generate accurately. For other characters such as Arknights Endfield Characters are a bit tricky because they also exist in the original Arknights. For example, since Laevatain is a variant of Surtr, you may need to include 'Surtr' in your prompt to help the model recognize her better.
Regarding your questions:
1. Training from zero: Since this is a personal project without sponsors, training from zero is too expensive for me.
2. Dataset size: In my experience, 100 images aren't quite enough for a character to be accurate, I usually need 400+ for the best results.
3. Technical suggestions: Training specific Unet blocks is an interesting idea, but it’s quite complex for full model training. Thanks for the input, though!
Thanks for your support and stay tuned for the update!
@Raelina No problem, I'm glad I could help a bit.
I can understand Nangong Yu not working as well, I was assuming 300ish would be enough for model training but as you stated you need 400+, which is an interesting detail. I have yet to try training a model outside Loras, so thank you for sharing that kind of detail.
Regarding Endfield, I was surprised to learn that too, especially because Danbooru has them as "character name"_(arknight), not even Endfield or Arknight Endfield, so dunno how that can confuse the model or even have an effect.
And thank you to answer my questions, I assume that was the case in training the database again compared to adding over. Just wanted to ask in case you have done it before.
Regarding the Dataset needing to be 400+ images for better representations, I noticed repeats aren't used (I only see Epochs been mentioned, unless I missed it). To my understanding (correct me if I'm wrong of course), repeats do repeat the dataset inside an Epoch, so each Epoch is like a step done in the training. Would having 200ish images per character and 2 or 3 repeats per epoch make the training more stable? This is just based on my Lora training and for stuff I have read, but I'm curious to know if this applies too to Full model training, I assume you have tried it before.
Either way, thanks again for indulging in my questions, so far it has been educative and entertaining to understand the process.
I was thinking to ask for some extra additions, but I noticed the characters don't have more than 300ish images, so it might not be worth asking for them atm. lol
So for now, I can't wait to see the changes you make for the next update, hopefully it works as planned!
@darionk I don't recommend using repeats for full model training. The dataset structure is quite different from what people usually use for Lora, and using repeats can cause the model to overfit much faster.
For more details on this dataset structure, I suggest reading the Animagine/Cagliostrolab blog or the Illustrious paper, as they go into great detail about the training process.
@Raelina Thank you for the extra info in your experience, I will give those papers a read. 😁
The background light of the image I generate will turn green and become very dark because it looks like this model is really awesome so I want to fix this problem
Thanks for the feedback! Could you please share the prompt and negative prompt you used? I’ll try to investigate what’s causing the issue
I am very happy that you can see my problem. It has been improved today. When I adjust the lora of CunnystyleV10 to 0.8, it will turn green. When I change the lora of other styles, some of them are normal and some will turn yellow but not very dark. There does not seem to be this problem when using ComfyUI. Because I am new to it, the pictures I create are not good enough and I do not continue to use ComfyUI. Here are the prompt words I use
Positive prompt words: sophia_claudel_kiyoubimbou,thighhighs,twintails,long hair,full body,black thighhighs,shoes,red hair,red footwear,ribbon,hair ornament,red eyes,anime coloring,<lora:XL-SophiaClaudel-KiyouBimbou-ILXL_r1:1>,[artist:yygg],
highres,extremely detailed 8k wallpaper,illustration,depth of field,great quality,masterpiece,best quality,amazing quality,very aesthetic,absurdres,detail eyes,
obese man,pillow,smartphone,cellphone,holding phone,panties around one leg,dark-skinned male,netorare,cheating (relationship),smaller domination,size difference,endure to moan,on back,on_side,quilt disappears,condoms filled with semen,rear insert,grabbing_from_behind,futon,tatami,faceless male,ugly man,nipples,hip_focus,bald,ugly bastard,twitching,splashing,lip_biting,endured_face,fully naked,<lora:CunnystyleV10:1>,
Reverse prompt words: lowres,jpeg artifacts,worst quality,watermark,very displeasing,low quality,bad anatomy disfigured malformed mutated,mutated hands and fingers,long body,bad anatomy,disfigured,malformed,three legs,bad hand,logo,Extra fingers,uncensored,negativeXL_D,word,signature,