tutu's HiSilk (Aurora 5D Black Pantyhose) | 图图的嗨丝（极光5D黑色连裤袜） | チュチュのハイシルク（オーロラ5D黒色タイツ）

Created with Tutu AI Workflow

Trained with Tutu Trainer. Tagged with Tutu Annotator.

This model is made inside my own Tutu AI workflow. Tutu Trainer is used for LoRA training, model resource management, parameter presets, and training checks. Tutu Annotator is used for dataset annotation, prompt reverse engineering, tag cleanup, English-friendly captions, and searchable tag organization.

I build these tools because LoRA quality is not only about pressing train. A good model needs cleaner captions, more stable training, careful tests, and a workflow that keeps the final model easy for other users to understand and reproduce. Tutu Trainer is free to use.

Tutu AI Workflow - Tutu Trainer and Tutu Annotator

Official website: https://zhaotutu.xyz/
Tutu Trainer: https://zhaotutu.xyz/downloads/tututrainer/
Tutu Annotator: https://zhaotutu.xyz/downloads/tutuannotator/
Bilibili: https://space.bilibili.com/431046154
YouTube: https://www.youtube.com/@zhaotutu/videos
Telegram: https://t.me/zhaotutu
TikTok: https://www.tiktok.com/@zhaotututu
Reddit: https://www.reddit.com/user/zhaotututu/
GitHub: https://github.com/zhaotututu
Email: [email protected]
QQ: 331506796

Tutu's HiSilk Aurora 5D Black Pantyhose LoRA

This is a realistic black pantyhose texture LoRA from my HiSilk series. The goal is very specific: improve the look of thin black pantyhose while keeping the rest of the image clean. It should not strongly pollute the face, shoes, clothing, pose, lighting, or overall color tone.

The latest version is v3.0. It is not just a small update from v2.0. I rebuilt the model around real-shot material, Difference Extraction Method 3.0, deeper dataset processing, and a new model-fusion test process. The visible improvements are lower pollution, stronger generalization, easier prompting, and better compatibility with different base models.

Brand / product reference: Moumi / 某觅
Series: Aurora Pure Enjoyment 2.0 / 极光纯享 2.0
Material: 8D 528N
Color: black
Base model: SD 1.5
Main trigger: tutututu, black pantyhose
Crotchless variant: add Crotchless pantyhose when needed

Why I made another black pantyhose LoRA

There are already many similar LoRAs, but a lot of them solve the texture problem by also dragging the rest of the picture along with it. Shoes may become wrong, clothing can be contaminated, the body can become stiff, or the whole picture may shift into an unwanted tone. Some models also mix different thicknesses and materials together, so the result becomes unstable.

This LoRA is designed for a narrower target. It focuses on thin glossy black pantyhose texture and tries to leave unrelated parts of the image alone. The small file size is not a weakness; it comes from the difference-extraction workflow. When the extraction is clean, the model can be lighter while still carrying the material feature it needs.

Recommended settings

Make sure the LoRA is loaded correctly in your WebUI or generation tool.
Set CLIP skip / CLIP stop layers to 2 if your workflow uses that setting.
For v3.0, start around 0.6 weight. You can usually test from 0 to 1 depending on the strength you want.
For old v1.0, the weight often needed to be higher, around 1 to 2. v3.0 is much easier to use.
Start CFG Scale around 7, then adjust with your base model and prompt style.
I often like DDIM, but other samplers can also work well. Please test with your own workflow.
Do not overload the negative prompt. Too many negatives can suppress the LoRA texture and reduce the effect.

Prompt keywords

Basic trigger:

tutututu, black pantyhose

Useful prompt phrases:

black pantyhose
thin black pantyhose
glossy black pantyhose
shiny black pantyhose
sheer black pantyhose
realistic hosiery texture
delicate nylon texture
subtle leg shine
natural light
Crotchless pantyhose

For best results, combine the trigger with a clear outfit, camera distance, pose, and lighting description. If the hosiery becomes too strong or starts affecting shoes and clothing, reduce LoRA weight first before adding more negative prompts.

Recommended LoRAs and models to combine

These were part of my original usage notes and are still useful as direction references. They are not mandatory, but they help if you want a polished portrait or fashion-photo result.

tutu's face / 图图的脸模 / 花ひよこ顔モデル
ADD DETAILS
OC
LOWRA
epiNoiseoffset
Film style / 胶片风

V3.0 production notes: science, art, and a lot of wasted stockings

V3.0 was made because v2.0 still had one obvious weakness: the texture improved, but the brightness and aurora feeling were not strong enough. To fix that, I did not only change training parameters. I went back to the source material.

I compared a number of real products, including Jirun K9, Jirun K1 2.0, Laimi Aurora Pure Enjoyment 2.0, Laimi Aurora, Laimi Dark Gold 2.0, Ling Liuguang, JM, and SK Aurora high-density. After testing, Laimi Aurora Pure Enjoyment 2.0 gave the best practical effect for this version.

Some tested products:

Laimi Aurora Pure Enjoyment 2.0:

Jirun K9:

Jirun K1 2.0:

Ling Liuguang:

The shooting method also changed. If all material is shot under the same studio light, the model may learn an unnatural center-line shine. For v3.0 I added natural-light material, so the final effect is less rigid and more usable in ordinary scenes.

During production, I used a self-developed fusion workflow and deeply processed the dataset. To avoid mismatch between training material and actual generation results, I created two LoRA sets first, then used intelligent fusion and exploration to form 50 candidate models.

After removing 25 weaker candidates, I compared the remaining 25:

The test focused on practical use, not only a single beautiful sample:

At weight 1, the face should not collapse even without face restoration.
Without extra clothing descriptors, the LoRA should affect black pantyhose as much as possible and other elements as little as possible.
Shoes, pose, and movement should stay stable.
The overall picture tone should not shift strongly.
The model should still work when used with several other LoRAs.
Different base models should remain friendly enough for normal users.

Models 17 and 24 entered the next round, and I finally selected No. 24.

This is why I say training is both science and art. The science is the dataset, difference analysis, fusion, comparison, and repeated testing. The art is knowing which result is actually pleasant to use. Making one hosiery LoRA even cost around 1000 RMB in total, with about 700 RMB spent just on products. Very unreasonable, very real, and somehow still worth doing.

Difference Extraction Method 3.0

Most LoRAs are trained directly from images. That works, but the model may also learn the base image style, lighting, background, outfit, shoes, and other unwanted factors. Difference extraction tries to preserve the useful difference while reducing those extra contaminations.

For this series, I optimized the process around my own material. It is not a simple textbook method. The workflow includes automated model exploration, LoRA fusion tests, and difference analysis based on Euclidean distance between LoRA weights. In simple terms, I treat weights as points in a high-dimensional space, compare how far two models differ, and use that difference to guide fusion decisions.

When two candidate models differ greatly in certain weights, those parts need special handling. When they are close, the fusion can be more conservative. This helps the final LoRA keep the hosiery texture while reducing unwanted pollution.

Version history

v3.0 - current recommended version

Rebuilt around real-shot material and Difference Extraction Method 3.0.
Uses Moumi / Laimi Aurora Pure Enjoyment 2.0 material as the main reference.
Added natural-light material to reduce unnatural studio-light artifacts.
Two LoRA sets were fused and explored into 50 candidate models.
Final candidate selected after comparing pollution, face stability, shoes, movement, tone shift, and base-model compatibility.
Trigger word unified to tutututu, black pantyhose for easier use.

v2.0

Reduced pollution to feet and shoes.
Changed the dataset to bf16 + 1024 real-shot material before extracting differences.
Reduced composition pollution and lowered the chance of short-limb issues.
Added the crotchless element. Use Crotchless pantyhose for that direction.
Main weakness: texture improved, but the aurora shine was still not bright enough.

v1.0

The first black pantyhose LoRA in this line.
Small file size due to the difference-extraction approach.
Made from real high-definition photos.
Had little impact on non-hosiery elements, but needed a higher weight, usually 1 to 2.
Reference product: Limer / Aurora / 5D / black.

FAQ

Why can I not generate the exact same image as your examples?

Different base models, WebUI versions, plugins, LoRA combinations, seeds, samplers, VAE, and hardware environments can all change the final image. Please do not chase an identical picture. Use the model to create your own best result.

What should I do if a large mosaic or broken texture appears after increasing the weight?

Lower this LoRA weight first. If you are using multiple LoRAs, also lower the weight of the other LoRAs. You can also try LoRA Block Weight or a different sampler/base model.

Why does the effect become weaker when I add many negative prompts?

Very heavy negative prompts may suppress the material features the LoRA is trying to add. Start with a clean negative prompt, then add only what the image actually needs.

Can I use it commercially?

You are responsible for the content you generate. If you use this model for commercial work or custom cooperation, please contact me first.

Statement

You are fully responsible for any creative work made with this model.
Do not use this model to create or share illegal, harmful, malicious, defamatory, fraudulent, or political content.
If you use this model commercially, please inform me. Thank you.

中文说明

这是图图嗨丝系列里的 极光 5D 黑色连裤袜 LoRA。它不是单纯让画面“更黑”或者“更亮”，而是尽量只强化黑色连裤袜的真实质感，同时减少对脸、鞋子、衣服、动作、构图和整体色调的污染。

当前推荐使用 v3.0。这一版相对 v2.0 不是小修小补，而是围绕实拍素材、差异提取法 3.0、数据集深度处理、两组 LoRA 融合探索、50 个候选模型测试重新做了一遍。重点提升是：污染更少、泛化更好、使用更方便、对不同底模更友好。

品牌 / 产品参考：某觅
所属系列：极光纯享 2.0
材质：8D 528N
颜色：black
基础模型：SD 1.5
触发词：tutututu, black pantyhose
开裆方向：需要时加入 Crotchless pantyhose

为什么还要再做一个黑丝 LoRA

同类 LoRA 很多，但很多模型会把黑丝质感和不该学的东西一起学进去，比如鞋子变形、衣服被污染、动作僵硬、画面整体偏色、厚度和材质混在一起。我的目标更窄：只尽量改善 5D 黑色连裤袜的薄、亮、透、细腻质感，把其他部分的影响降到最低。

这个 LoRA 体积小不是偷工减料，而是差异提取工作流带来的结果。提取越干净，模型越不需要把背景、灯光、底模风格等东西也一起塞进去。

出图关键词

基础触发：tutututu, black pantyhose

black pantyhose
thin black pantyhose
glossy black pantyhose
shiny black pantyhose
sheer black pantyhose
realistic hosiery texture
delicate nylon texture
subtle leg shine
natural light
Crotchless pantyhose

如果出现鞋子、衣服、肤色被带偏的情况，优先降低 LoRA 权重，不要一上来就堆负面词。这个模型本来就是为了减少污染做的，干净的参数往往更容易看到它的优势。

制作过程：科学、审美和一点点离谱

v2.0 的问题主要是不够亮，虽然质感已经提高，也有极光的感觉，但还没有达到我想要的状态。所以 v3.0 不是只改训练参数，而是重新从素材端开始解决。

我购买并对比了多款产品：极润 K9、极润 K1 2.0、莱觅极光纯享 2.0、莱觅极光、莱觅暗金 2.0、凌琉光、JM 妙不可言、SK 极光高密版等。最终测试后，莱觅极光纯享 2.0 的实际效果最好。

拍摄方式也做了调整。前两个版本有时会因为影棚光太统一，学到中间一条线一样的高光。v3.0 加入了自然光素材，让效果更自然，也更适合普通场景。

制作上，我使用了自研融合流程，对数据集做了深度处理。为了避免素材效果和实际出图不匹配，我先做了两套 LoRA，再通过智能融合和探索形成 50 个候选模型。排除 25 个弱样本后，对剩余候选进行横向对比，重点看权重 1 时脸会不会崩、鞋子会不会污染、动作会不会受影响、整体色调会不会漂移、多 LoRA 混用时是否稳定。最后选择了 24 号。

所以这个模型看起来只是一个黑色连裤袜 LoRA，背后其实是一次很费钱也很费劲的测试。做一个模型倒贴接近一千块，买嗨丝就花了七百块。一地鸡毛，但模型能更好用，这事也算值了。

差异提取法 3.0

普通 LoRA 直接用图片训练，很容易把底图风格、灯光、背景、衣服、鞋子等不相关内容一起学进去。差异提取法的目标，是尽量把真正有用的差异提取出来，把这些污染减掉。

这一版我针对自己的素材做了进一步优化，不是传统意义上简单套流程。包括自动化模型探索、LoRA 融合测试，以及基于欧氏距离的权重差异分析。简单说，就是把 LoRA 权重看成高维空间里的点，比较模型之间的距离，再根据差异决定融合方式。差异大的权重需要特别处理，差异小的部分可以更保守地融合。

版本说明

v3.0：当前推荐版本。 实拍素材重做，差异提取法 3.0，加入自然光素材，两套 LoRA 融合探索出 50 个候选模型，最终根据污染、脸部稳定、鞋子、动作、色调、底模兼容性选出 24 号。触发词统一为 tutututu, black pantyhose。

v2.0： 降低对脚和鞋子的污染，数据集改为 bf16 + 1024 实拍后再做差异提取，构图污染更少，短肢等问题更少，并加入开裆元素。缺点是亮度和极光感还不够，所以 v3.0 主要解决这个问题。

v1.0： 本系列最早版本，使用差异提取法和高清实拍素材制作，体积小，对画面其他内容影响较少，但需要较高权重，通常 1 到 2 效果更明显。参考产品为极光 5D 黑色款。

常见问题

为什么我出不了和你一样的图？

底模、插件、WebUI 版本、VAE、采样器、种子、LoRA 组合、显卡环境都会影响结果。不要追求完全一样的图，做出适合你自己画面的效果更重要。

权重调高后出现马赛克或破碎纹理怎么办？

先降低本 LoRA 权重。如果你同时用了很多 LoRA，也降低其他 LoRA 的权重，或者尝试 LoRA Block Weight、换采样器、换底模。

为什么负面提示词太多反而效果不好？

过重的负面词可能把材质特征也压掉。建议先用干净的负面词，看到问题后再逐步补。

声明

你需要对使用本模型生成的任何作品承担全部责任。
不要使用本模型制作或传播非法、有害、恶意、诽谤、诈骗或政治用途内容。
如果将本模型用于商业用途或定制合作，请提前联系我，谢谢。

日本語説明

これは Tutu HiSilk シリーズの Aurora 5D Black Pantyhose LoRA です。目的は、薄い黒タイツのリアルな質感を強めながら、顔、靴、服、ポーズ、画面全体の色調への影響をできるだけ小さくすることです。

おすすめは v3.0 です。このバージョンは単なる微調整ではなく、実写素材、Difference Extraction Method 3.0、データセット処理、2 つの LoRA セットの融合、50 個の候補モデル比較を通して作り直したものです。

Reference product: Moumi / Aurora Pure Enjoyment 2.0
Material: 8D 528N
Color: black
Base model: SD 1.5
Trigger: tutututu, black pantyhose
Crotchless style: add Crotchless pantyhose

v3.0 の推奨ウェイトはまず 0.6 付近から試してください。CLIP skip は 2、CFG Scale は 7 付近から調整できます。ネガティブプロンプトを入れすぎると質感が弱くなる場合があります。

v2.0 は靴や足への汚染を減らし、bf16 + 1024 の実写素材から差分抽出を行ったバージョンです。v1.0 は最初の 5D 黒タイツ版で、小さなファイルサイズと低い汚染が特徴ですが、より高いウェイトが必要でした。

この LoRA は、黒タイツの質感だけをなるべく自然に追加するためのモデルです。同じ画像を再現するより、自分のベースモデルとワークフローに合わせて調整してください。

Model Type	LORA
Base Model	SD 1.5
Published	2024-01-25
Trained Words	tutututu,blackpantyhose,