モデルカード / Model Card
モデル名 / Model Name: yoiko_style_model
親モデル / Parent Model: WAI-NSFW-illustrious-SDXL Ver15.0
ライセンス / License: Illustrious License
学習データ / Training Data: 私が生成した画像（未公開を含む） / Images generated by me (including some not publicly released)
紹介文（日本語）
本モデルは WAI-NSFW-illustrious-SDXL Ver15.0 をベースに、追加の SFT（Supervised Fine-Tuning） を施した公開用ベースモデルです。絵柄は過去の yoiko_lora シリーズに近い傾向があります。これは、これまでの学習データと使用画像が類似しているためで、yoiko_lora シリーズが持つ低レベル表現の重みをより丁寧に更新することを目指しました。
ベースとして WAI-NSFW-illustrious-SDXL Ver15.0 を選んだ理由は、圧倒的な完成度と、noob ai 系と比べてライセンスが比較的緩やかで試行錯誤の幅が広い点にあります。現時点でも一定の完成度はあると考えていますが、生成画像の品質には未知数な部分が残っています。ぜひフィードバックをお寄せください。
それでは、楽しんでご利用ください。
Description (English)
This public base model builds upon WAI-NSFW-illustrious-SDXL Ver15.0 with additional Supervised Fine-Tuning (SFT). Its visual style is close to my previous yoiko_lora series. That’s because the training and reference images are similar; the goal here is to refine and update the lower-level representational weights that characterize the yoiko_lora line.
I chose WAI-NSFW-illustrious-SDXL Ver15.0 due to its excellent baseline quality and a more permissive license (compared to some noob-ai variants), which allows broader experimentation. While I believe the model is reasonably polished, output quality may still vary, and there are unknowns to explore. Feedback is very welcome.
In short—have fun creating!
Description
JP
本モデルは WAI-NSFW-illustrious-SDXL Ver15.0 をベースに SFT（Supervised Fine-Tuning） を行った 初版 です。
当初の学習過程で過学習の傾向が見られたため、予定より step 数を少し削減して最終化しています。絵柄は過去の yoiko_lora シリーズに近く、これは学習・使用画像が類似しているためです。目標は、yoiko_lora シリーズの低レベル表現の重みを更新・洗練することでした。
ベースとして WAI-NSFW-illustrious-SDXL Ver15.0 を選んだ理由は、完成度の高さと、noob ai 系と比べて相対的に緩やかなライセンスで試行錯誤の幅が広い点です。完成度は一定レベルにありますが、生成品質には未知数な部分も残っています。ぜひフィードバックをお願いします。
それでは、楽しんでご利用ください。
ENG
This is the first release of a public base model fine-tuned (SFT) on top of WAI-NSFW-illustrious-SDXL Ver15.0.
During training, we observed signs of overfitting, so the final step count was slightly reduced from the original plan. The visual style is close to my yoiko_lora series because the training and reference images are similar. The aim was to refine and update the lower-level representational weights characteristic of the yoiko_lora line.
WAI-NSFW-illustrious-SDXL Ver15.0 was chosen for its strong baseline quality and a more permissive license (relative to some noob-ai variants), which enables broader experimentation. While reasonably polished, output quality may still vary and there are unknowns to explore. Feedback is highly appreciated.
In short—have fun creating!
FAQ
Comments (5)
bounty‼️
→https://civitai.com/bounties/9945/yoikostylemodels-publish-test-bounty until 2025/10/25
I accidentally found that this checkpoint works surprisingly well with my custom LoRA.
However, it does not perform as well on word combination tagging as Danbooru's existing tags.
(e.g., grabbing another's pussy, guided pussy grab)
Also, I've noticed it struggles with generating images where hair covers an eye, as the covered eye is often distorted or poorly rendered.
(e.g, side-swept bangs + one side hair fringe over one eye)
update: it struggles at generating feet too.
And I hope that hand generation can become more stable to reduce the need for rerolling due to instability in hand generation.
For the foreseeable future, I'll probably be using this as my main checkpoint.
thx for feedbacking.Thank you for your feedback.
I’m currently working on adjusting and improving the efficiency of the annotation process.
Previously, I was only doing simple annotations, but now I’m trying to add more detailed information.
As a result, I expect this improvement will soon have a positive impact on LoRA training.
You also gave me feedback before about model forgetting.
However, it’s true that images where hair covers the eyes are either very few or maybe not included at all in the training data.
That was a complete blind spot for me.
It’s really hard to eliminate bias when working alone.
Please also try the other checkpoint, Kadenz, if you have a chance.
It might also struggle with hiding the eyes, though...
Once again, thank you so much for your valuable feedback.
@ARMOR_AI
Maybe it's because of your simple annotations, When I generate images, this checkpoint sometimes produces unexpected surprises (e.g., clothing, backgrounds...etc).
The high frequency of these surprises is the main reason I enjoy using this checkpoint. I suspect this effect might be a result of "prompt contamination?"
As you work on improving accuracy, I hope the model can still retain this ability to generate these surprises, rather than becoming so precise that the resulting images lose that creative element.
@Pavers That term prompt contamination is new to me.
I looked into it a bit, and it reminded me of the reason why I first fell in love with AI-generated images — I had completely forgotten about that.
I can’t say for sure what direction things will go, but maybe those unexpected, spontaneous expressions are what make AI art so fascinating.
I still have the pre-annotation data, so I think I managed to refine it while keeping that sense of “spontaneity.”
It might be difficult to balance precision and unpredictability, but I’ll keep experimenting!
AI is really challenging, isn’t it?
P.S. Sorry if I misunderstood anything.