V2.0 版发布说明V2.0 核心改进： 优化了打标规则，解决了 V1.0 触发词难以带起 LoRA 的痛点。触发词用法：V2.0 的触发词需包含下划线，以确保不会被语义稀释。推荐格式：opaque_tights, black_matte_tights,建议把触发词放在提示词开头，后续自然语言描述中简单提一下厚裤袜即可。兼容性<ul><li>可一定程度叠加 NSFW LoRA，叠加后裤袜效果可能会受影响（因为没有找到厚裤袜的 NSFW 图片做训练集）</li></ul>权重建议推荐权重范围 0 ~ 2：权重效果0-2。大于2可能丢失颜色，只保留材质效果💡 小技巧： 利用颜色丢失的特性，不可以写下划线触发词的话，可以做出类似油光细腻皮肤的效果。如果裤袜出现缝合线的话，需要在提示词中加注没有缝合线的相关提示词---V2.0 Release NotesV2.0 Key Improvement: Refined the captioning rules, solving V1.0's issue where trigger words struggled to activate the LoRA.Trigger Word Usage:V2.0 trigger words must include underscores to prevent semantic dilution.Recommended format: opaque_tights, black_matte_tights,It's recommended to place trigger words at the beginning of the prompt, and briefly mention opaque tights in the subsequent natural language description.CompatibilityCan be combined with NSFW LoRAs to a certain extent. The tights effect may be affected when combined (because no NSFW images of opaque tights were found to use as training data).Weight RecommendationRecommended weight range: 0 ~ 2:Weight effect 0-2. Above 2 may lose color, retaining only the material effect.💡 Tip: Take advantage of the color-loss characteristic. If you cannot use the underscore trigger word, you can achieve a glossy smooth skin-like effect. If visible seam lines appear on the tights, add relevant negative prompts about no seam lines to your prompt.<hr /><hr />V1 — Krea2-pantyhose-V3_B-000004（当前发布版）训练日期： 2026-07-12 数据集： 116 张，4 色 × 2 光泽，不透明 打标： 自然语言英文句子 训练框架： Fizgig（Krea 2 RAW） 推荐检查点： 第 4 轮（loss 0.0656） 备用检查点： 第 3 轮（loss 0.0689）<ul><li>黑色效果最稳，高权重下依然保持厚实质感</li><li>白/灰/肉色高权重时也能出厚裤袜效果，但不如黑色稳定</li><li>权重范围 0~3，越高越厚实不透肉</li><li>底模必须是 Krea 2 Turbo，8 步推理</li><li>后续计划：如果大家喜欢的话可以给我点赞或者留言，我会尝试去优化这个lora</li></ul>V1 — Krea2-pantyhose-V3_B-000004 (Current Release)Training date: 2026-07-12 Dataset: 116 images, 4 colors × 2 finishes, opaque only Captions: Natural language English sentences Training framework: Fizgig (Krea 2 RAW) Recommended checkpoint: Epoch 4 (loss 0.0656) Alternative checkpoint: Epoch 3 (loss 0.0689)<ul><li>Black performs best — stays thick and solid even at high weights</li><li>White/grey/nude can also achieve thick opaque results at higher weights, but less consistently than black</li><li>Weight range: 0–3, higher = thicker and more opaque</li><li>Base model must be Krea 2 Turbo, 8-step inference</li><li>Trigger word: <code>pantyhose</code></li><li>Future plans: If you like this LoRA, feel free to leave a like or comment. I'll keep improving it based on feedback.</li></ul><h3>背景</h3>在 C 站寻找 Krea 2 的裤袜 LoRA 时，我发现已有的裤袜 LoRA 虽然效果不错，但在厚裤袜（不透明）的表现上稍微没有那么给力。所以开始自己学习如何训练 LoRA、如何选择训练集以及如何打标。这个 LoRA 是我第一个发布的 LoRA，可能有很多不足之处，也希望各位能给我指出和建议。<h3>功能介绍</h3><ul><li>不透明材质 — 厚实不透肉，颜色饱满均匀</li><li>哑光 vs 油光 — 两种光泽效果清晰可辨</li><li>4 种基础颜色：黑色、白色、灰色、肉色</li><li>细节丰富 — 脚底防滑条、面料褶皱、紧身包裹感一应俱全</li></ul><h3>使用方式</h3><ul><li>底模： Krea 2 Turbo（训练基于 Krea 2 RAW）</li><li>触发词： <code>pantyhose</code></li><li>推荐权重： 0–3.0（权重越高裤袜越厚实不透肉）</li><li>步数： 8 步（Turbo）</li></ul>注意： 黑色效果最佳，高权重下依然保持厚实哑光/油光质感。其他颜色（白/灰/肉色）在高权重时也能达到厚裤袜效果，但不如黑色稳定。<h3>附加效果</h3>除了控制厚度和光泽，这款 LoRA 还能改善裤袜的纹理细节——包括面料光泽感、编织纹理和腿部缝合线的真实感，让裤袜看起来更有材质质感而非简单的皮肤上色。<h3 id="8">提示词参考 — 8 种组合</h3>组合关键词黑色哑光不透明black matte opaque黑色油光不透明black glossy opaque灰色哑光不透明grey matte opaque灰色油光不透明grey glossy opaque白色哑光不透明white matte opaque白色油光不透明white glossy opaque肉色哑光不透明nude matte opaque肉色油光不透明nude glossy opaque<h3>训练参数</h3><ul><li>数据集：116 张，4 色 × 2 光泽</li><li>打标方式：自然语言句子</li><li>分辨率：1024×1024</li><li>训练框架：Fizgig（Krea 2 RAW）</li></ul><hr /><h2 id="keea2-opaque-pantyhose-lora">Keea2 Opaque Pantyhose LoRA</h2>A Krea 2 LoRA trained exclusively on opaque pantyhose in 4 colors × 2 finishes, giving you precise control over hosiery in your generations.<h3 id="backstory">Backstory</h3>When I searched for Krea 2 pantyhose LoRAs on Civitai, I found some great options, but most of them didn't perform as well on thick/opaque pantyhose specifically. So I started learning how to train LoRAs — how to curate datasets, how to caption them properly, and how to tune the training process. This is my very first published LoRA, and there's definitely room for improvement. Any feedback and suggestions are welcome!<h3 id="features">Features</h3><ul><li>Opaque only — Solid color coverage, non-transparent</li><li>Matte vs Glossy — Two finishes clearly distinguishable</li><li>4 base colors: black, white, grey, nude</li><li>Rich details — Grip soles, fabric folds, skin-tight fit</li></ul><h3 id="how-to-use">How to Use</h3><ul><li>Base model: Krea 2 Turbo (trained on Krea 2 RAW)</li><li>Trigger word: <code>pantyhose</code></li><li>Weight range: 0–3.0 (higher = thicker, more opaque)</li><li>Steps: 8 (Turbo)</li></ul>Note: Black works best at high weights. Other colors (white/grey/nude) can achieve thick opaque results at higher weights but less consistently than black.<h3 id="additional-effects">Additional Effects</h3>Beyond opacity and gloss control, this LoRA improves the texture details of pantyhose — fabric sheen, knit texture, and seam lines — making hosiery look like real fabric rather than painted-on skin.<h3 id="prompt-reference-8-combinations">Prompt Reference — 8 Combinations</h3>LookKeywordsBlack matte opaqueblack matte opaqueBlack glossy opaqueblack glossy opaqueGrey matte opaquegrey matte opaqueGrey glossy opaquegrey glossy opaqueWhite matte opaquewhite matte opaqueWhite glossy opaquewhite glossy opaqueNude matte opaquenude matte opaqueNude glossy opaquenude glossy opaque<h3 id="training-details">Training Details</h3><ul><li>Dataset: 116 images, 4 colors × 2 finishes</li><li>Captions: Natural language sentences</li><li>Resolution: 1024×1024</li><li>Framework: Fizgig (Krea 2 RAW)</li></ul>

Krea2 Opaque Pantyhose 厚裤袜

V1 — Keea2-pantyhose-V3_B-000004（当前发布版）训练日期： 2026-07-12 数据集： 116 张，4 色 × 2 光泽，不透明 打标： 自然语言英文句子 训练框架： Fizgig（Krea 2 RAW） 推荐检查点： 第 4 轮（loss 0.0656） 备用检查点： 第 3 轮（loss 0.0689）<ul><li>黑色效果最稳，高权重下依然保持厚实质感</li><li>白/灰/肉色高权重时也能出厚裤袜效果，但不如黑色稳定</li><li>权重范围 0~3，越高越厚实不透肉</li><li>底模必须是 Krea 2 Turbo，8 步推理</li><li>后续计划：如果大家喜欢的话可以给我点赞或者留言，我会尝试去优化这个lora</li></ul>V1 — Keea2-pantyhose-V3_B-000004 (Current Release)Training date: 2026-07-12 Dataset: 116 images, 4 colors × 2 finishes, opaque only Captions: Natural language English sentences Training framework: Fizgig (Krea 2 RAW) Recommended checkpoint: Epoch 4 (loss 0.0656) Alternative checkpoint: Epoch 3 (loss 0.0689)<ul><li>Black performs best — stays thick and solid even at high weights</li><li>White/grey/nude can also achieve thick opaque results at higher weights, but less consistently than black</li><li>Weight range: 0–3, higher = thicker and more opaque</li><li>Base model must be Krea 2 Turbo, 8-step inference</li><li>Trigger word: <code>pantyhose</code></li><li>Future plans: If you like this LoRA, feel free to leave a like or comment. I'll keep improving it based on feedback.</li></ul>

Files with SHA256: 6bf92dc33056bc792d3379a9e690f54d652bda25496dbf6823308a1a66dae0f3