其实moss的底模对厚裤袜是有训练的,可以通过提示词去唤出,这让我很惊喜。但是,但是,但是偶尔还是不那么容易去琢磨与磨合这个提示词,所以我还是决定训练moss的厚裤袜lora,那么,欢迎大家使用。
中/英双版:
english:
# MOSS-Image — Opaque Tights & Pantyhose (10 Colour × Finish Combos)
Base model: MOSS-Image (trained on MOSS-Image Base, best used on SFT)
A trigger-driven LoRA for opaque tights and pantyhose, built for MOSS-Image. Ten colour × finish combinations, so the tights can be matched to the light and the outfit instead of defaulting to black.
## Trigger words
opaque_tights is the base trigger — add exactly one colour/finish pair from the table below.
| Trigger combination | Colour | Finish | Images | Suits |
|---|---|---|---|---|
| opaque_tights, black_matte_tights | black | matte | 24 | dim rooms, warm indoor light, dark outfits |
| opaque_tights, white_matte_tights | white | matte | 17 | bright, airy scenes |
| opaque_tights, flesh_matte_tights | flesh | matte | 16 | soft daylight, natural look |
| opaque_tights, flesh_glossy_tights | flesh | glossy | 15 | hard light, sheen-focused shots |
| opaque_tights, black_glossy_tights | black | glossy | 15 | hard directional light, night scenes |
| opaque_tights, nude_matte_tights | nude | matte | 15 | soft light, everyday natural look |
| opaque_tights, white_glossy_tights | white | glossy | 12 | hard light against a darker set |
| opaque_tights, gray_matte_tights | grey | matte | 11 | bright, low-saturation daylight |
| opaque_tights, gray_glossy_tights | grey | glossy | 10 | hard side light |
| opaque_tights, dark_blue_matte_tights | dark blue | matte | 9 | cool-toned scenes, dark outfits |
Note: in the dataset, nude and dark_blue only exist in the matte finish; black / white / flesh / grey come in both matte and glossy.
## How to prompt it
1. Put the triggers at the very front. It works best when opaque_tights, <colour>_<finish>_tights sits in the first comma segment, ahead of the rest of the sentence.
2. Relatively speaking, skipping the acceleration LoRA (lightning / turbo / distilled accelerator) tends to give thicker tights; run the full 50 steps.
3. If you are not satisfied with the result, you can compensate with the prompt: switch the wording to the opaque_tights system and avoid sheer-prior words such as nylon / sheer / pantyhose - write the gloss as oily high-gloss, never greasy, e.g. deep opaque black tights with an oily high-gloss sheen covering the legs fully; for the layout branch, write the tights sentence into summary at the very front AND again into a separate leg object box (double-write) - this is the most reliable method in testing. Place this before the character description - the earlier it appears, the stronger the control over the fabric.
4. Match the finish to the light. Matte reads under any light. Glossy needs a directional source (hard sun, side light or rim light); in fully diffused light the sheen will not come through.
5. LoRA weight: 0.9 recommended — you can adjust it yourself between 0 and 1.
6. Recommended setup: base model MOSS-Image; default sampler settings from your own MOSS workflow are fine.
## Training details
- Dataset: 144 images, training resolution 992 x 992 (buckets enabled; the weight metadata records 1024 x 1024)
- Network: LoRA dim 32 / alpha 32, 756 tensors, standard kohya format
- Training: 35 epochs, AdamW, lr 1e-4, base model MOSS-Image Base (bf16)
- Coverage: seated, standing, walking, bending and top-down leg close-ups; everyday and indoor/outdoor settings
中文:
# MOSS-Image — 不透明连裤袜 LoRA(10 组 颜色 × 光泽)
底模:MOSS-Image(训练基于 MOSS-Image Base,建议在 SFT 上使用)
面向 MOSS-Image 的不透明连裤袜触发词 LoRA,共 10 组颜色 × 光泽组合,让裤袜跟着光线与穿搭走,而不是每次都默认黑丝。
## 触发词
opaque_tights 是基础触发词,后面只加一组 颜色_光泽 组合(见下表)。
| 触发词组合 | 颜色 | 光泽 | 训练图数 | 适配场景 |
|---|---|---|---|---|
| opaque_tights, black_matte_tights | 黑 | 哑光 | 24 | 暗调房间、暖室内光、深色穿搭 |
| opaque_tights, white_matte_tights | 白 | 哑光 | 17 | 明亮通透的场景 |
| opaque_tights, flesh_matte_tights | 肉色 | 哑光 | 16 | 柔光日光、自然感 |
| opaque_tights, flesh_glossy_tights | 肉色 | 油亮 | 15 | 硬光、强调反光质感 |
| opaque_tights, black_glossy_tights | 黑 | 油亮 | 15 | 硬质定向光、夜景 |
| opaque_tights, nude_matte_tights | 裸色 | 哑光 | 15 | 柔光、日常自然 |
| opaque_tights, white_glossy_tights | 白 | 油亮 | 12 | 深色场景中的硬光 |
| opaque_tights, gray_matte_tights | 灰 | 哑光 | 11 | 明亮低饱和日光 |
| opaque_tights, gray_glossy_tights | 灰 | 油亮 | 10 | 硬质侧光 |
| opaque_tights, dark_blue_matte_tights | 深蓝 | 哑光 | 9 | 冷调场景、深色穿搭 |
说明:nude 与 dark_blue 在训练集中只有哑光档;black / white / flesh / gray 四色哑光与油亮齐备。
## 写法
1. 触发词放最前——opaque_tights, <颜色>_<光泽>_tights 落在第一个逗号段,排在句子其余部分之前。
2. 相对而言,不挂加速 LoRA(lightning / turbo / 蒸馏加速模型)更容易出厚裤袜;采样步数跑满 50 步。
3. 如果效果不满意,可以通过提示词补足:用词改用 opaque_tights 体系、避开 nylon / sheer / pantyhose 这类薄丝先验词——油亮写 oily high-gloss,不要写 greasy,示例:deep opaque black tights with an oily high-gloss sheen covering the legs fully;layout 支把裤袜句同时写进 summary 最前与一个独立的腿部 object 框(双写)——这是实测最稳的写法。这段放在人物描述之前——越靠前的描述,对袜料的控制越强。
4. 光泽档要配定向光:哑光任何光都能出;油亮需要硬质定向光(正射硬光/侧光/轮廓光),全漫射环境下光泽出不来。
5. LoRA 强度:推荐 0.9,0 – 1 之间可以自己调整。
6. 推荐配置:底模 MOSS-Image;采样参数沿用你自己的 MOSS 工作流默认值。
## 训练信息
- 数据集:144 张,训练分辨率 992 × 992(开启 bucket;权重元数据内标注 1024 × 1024)
- 网络:LoRA dim 32 / alpha 32,756 个张量,kohya 标准格式
- 训练:35 epoch · AdamW · 学习率 1e-4 · 底模 MOSS-Image Base(bf16)
- 覆盖:坐姿、站姿、行走、弯腰、俯拍腿部特写;日常实景与室内外环境
Description
V1.0 当然,是moss版本的1.0








