<h2 id="for-all-the-lora-lab-tests'-info.-()-radd57802">For all the LoRa Lab tests' info. (如果像了解更多轮次的尝试)</h2><h3 id="please-go-to-https:civitai.commodels1366934-cc9t7vrg4">please go to <a target="_blank" rel="ugc" href="https://civarchive.com/models/1366934">https://civarchive.com/models/1366934</a> </h3><h3 id="https:civitai.commodels1366934-llqugnorf">请移步 <a target="_blank" rel="ugc" href="https://civarchive.com/models/1366934">https://civarchive.com/models/1366934</a></h3><h2 id="-l43tc5bp8"></h2><h3 id="explanation-different-approach-of-current-test-suite-(-)-hs48yu6hq">Explanation different approach of current Test Suite (本轮训练实验集中, 不同方案的说明)</h3>Try the following approaches to see how the trained "nsfw stylized (pantyhose &amp; see-through)" LoRa feels and what the differences are:<ul><li>1 image, no regularization (p1_wor) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360124">https://civarchive.com/models/1360124</a></li><li>1 image, with regularization (p1_wr) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360141">https://civarchive.com/models/1360141</a></li><li>3 images, no regularization (p3_wor) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360163">https://civarchive.com/models/1360163</a></li><li>3 images, with regularization (p3_wr) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360171">https://civarchive.com/models/1360171</a></li><li>8 images, no regularization (p8_wor) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360183">https://civarchive.com/models/1360183</a></li><li>8 images, with regularization (p8_wr) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360189">https://civarchive.com/models/1360189</a></li><li>24 images, no regularization (p24_wor) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360230">https://civarchive.com/models/1360230</a></li><li>24 images, with regularization (p24_wr) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360268">https://civarchive.com/models/1360268</a></li></ul>This time captions are directly derived using WD14 Conv v2 tagger, without any additions or deletions尝试一下如下的方案, 所训练出来的 "nsfw 风格化 (pantyhose &amp; see-through)" 的 LoRa 感觉如何, 各自有何差别 :<ul><li>1 张图片, 无正则化 (p1_wor) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360124">https://civarchive.com/models/1360124</a></li><li>1 张图片, 有正则化 (p1_wr) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360141">https://civarchive.com/models/1360141</a></li><li>3 张图片, 无正则化 (p3_wor) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360163">https://civarchive.com/models/1360163</a></li><li>3 张图片, 有正则化 (p3_wr) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360171">https://civarchive.com/models/1360171</a></li><li>8 张图片, 无正则化 (p8_wor) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360183">https://civarchive.com/models/1360183</a></li><li>8 张图片, 有正则化 (p8_wr) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360189">https://civarchive.com/models/1360189</a></li><li>24 张图片, 无正则化 (p24_wor) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360230">https://civarchive.com/models/1360230</a></li><li>24 张图片, 有正则化 (p24_wr) : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1360268">https://civarchive.com/models/1360268</a></li></ul>本次 captions 尝试直接用 WD14 Conv v2 tagger 推导, 无任何的增删<h3 id="desired-training-effects-()-sxfoq4mvg">Desired Training Effects (训练想要的效果)</h3><ul><li>Increase the maturity of girls</li><li>Add natural transparency effect to clothing</li><li>Adjust the style of pantyhose (flesh-colored shiny style; and increase the default probability of pantyhose when not specified)</li><li>提升女孩的成熟度</li><li>加入衣物的原生透明效果</li><li>调整连裤袜的风格 (肉色舍宾风格; 且增加无指定时裤袜的默认几率)</li></ul><h2 id="some-findings-()-9fy03es4c">Some Findings (一些发现)</h2><h3 id="lab@1-250316-record-(lab@1-250316-)-llmaaagh2">Lab@1 - 250316 Record (Lab@1 - 250316 记录)</h3><ul><li>Some input images (dataset) and regularization images mentioned here can be referenced in the model's image showcase (showcase)<ul><li>One great thing about civit is that it displays image-related prompts and parameters very well, including XY comparison charts generated by webui</li></ul></li><li>p1 wor can actually pixel-level replicate the input girl's appearance in many scenes<ul><li>But the disadvantage is also very obvious: without special prompts, it tries very hard to restore all features, with little variability</li></ul></li><li><code>full body</code> in WAI's base model has a high coupling with <code>standing</code> posture<ul><li>But trying to add <code>standing</code> in neg prompt doesn't seem to have obvious effect</li></ul></li><li>The feeling of p24 training is that it does migrate girl's features to other characters, as seen in the test image of Ao Run (敖闰) where the dragon's face shows obvious feminization influence</li><li>p24 wor version, when +pantyhose is used, sometimes seems to produce rather peculiar poses (not sure if it's because most poses in the dataset are quite large)<ul><li>For details not learned (like hands, small objects, some detail logic, ...), p24 doesn't seem to handle them very well</li></ul></li><li>p24 wr feels visually the best in quality<ul><li>Although many features are not directly reflected, but rather have their own understanding and fusion</li><li>But its control ability over +full body seems problematic<ul><li>Not sure if it's related to reg images containing many partial body images</li></ul></li><li>Also, when pantyhose appears without other prompt controls, it tends to output lower body images</li></ul></li><li>Additionally, when dealing with complex compositions, p24 series sometimes tends to express more exaggeratedly, leading to image collapse</li><li>这里说的一些输入图片 (dataset), 还有正则化图片, 可以参考本个模型里面的图片演示 (showcase)<ul><li>civit 非常不错的一个点是, 它会把图片相关的提示词和参数很好的展示出来, 包括 webui 产出的 XY 对比图</li></ul></li><li>p1 wor 居然能在不少的场景下, 像素级别地复刻输入的女孩形象<ul><li>但劣势也非常明显, 在无特殊的提示词时, 它会非常努力地还原所有特征, 可变性不强</li></ul></li><li><code>full body</code> 在 WAI 这个底模 (base model) 里, 和 <code>standing</code> 的姿态有很高的耦合性<ul><li>而且, 尝试过在 neg prompt 里加上 <code>standing</code> 好像效果也不明显</li></ul></li><li>p24 训练下的感觉, 确实会将女孩的特征迁移到其他形象上, 如测试中的敖闰 (ao run) 图片中的 "龙" 的脸部有很明显的女性化影响</li><li>p24 wor 版本, 在 +pantyhose 时, 有时好像容易出现比较奇特的姿势 (不知道是否和数据集大多数姿势都比较大的原因)<ul><li>没有学习过的细节处理 (如手部, 一些小物件, 一些细节逻辑, ...), p24 好像处理起来不是太好</li></ul></li><li>p24 wr 的质量在直观上感觉最优<ul><li>虽然很多特征并没有很直接的体现, 而是有它自己的理解和融合</li><li>但是它对 +full body 的控制能力好像有问题<ul><li>不知道是否和 reg 图片包含很多局部身体图片有关</li></ul></li><li>并且, 一旦出现 pantyhose 且没有其他提示词控制时, 它倾向于输出 lower body 的图片</li></ul></li><li>另外, p24 系列处理复杂构图时, 有时会倾向于更夸张地去表达, 导致崩图</li></ul><h3 id="potential-issues-()-d1t1u2g90">Potential Issues (可能的问题)</h3><ul><li>The entire dataset is 1girl, solo, so the default style for other characters may have problems (such as male characters)</li><li>Not sure about LoRa's ability to handle complex prompts</li><li>全数据集都是 1girl, solo, 所以其他形象的默认风格可能会出现问题 (如男人形象)</li><li>不太确定 LoRa 在处理复杂 prompt 的能力</li></ul><h2 id="current-training-settings-()-1rqf4ajyv">Current Training Settings (本次训练相关设置)</h2><pre><code>Number of images: number (refer to p1, p3, p8, p24)
Regularization: with reg (wr) | without reg (wor)
</code></pre><pre><code>图片数量: number (参考 p1, p3, p8, p24)
是否正则化: with reg (wr) | without reg (wor)
</code></pre><h2 id="model-training-parameters-()-2ruq1tmjz">Model Training Parameters (模型训练参数)</h2><pre><code>[Training Parameters]
BASE=https://civarchive.com/models/827184/wai-nsfw-illustrious-sdxl =&gt; v12
EPOCHS=8 # Number of training epochs
SEED=1861

[Resolution Settings]
TRAINING_WIDTH=1024 # Training width
TRAINING_HEIGHT=1152 # Training height
MAX_BUCKET_RESO=$TRAINING_HEIGHT # Maximum bucket resolution

[Hardware Adaptation]
DIM=32 # Network dimension
DIM_ALPHA=28 # Network alpha value
TRAINING_BATCH_SIZE=1 # Batch size

[Training Parameters]
--train_batch_size 1
--learning_rate 0.0001
--unet_lr 0.0001
--text_encoder_lr 0.00001
--optimizer_type "AdamW8bit"
--mixed_precision "bf16"
--save_model_as "safetensors"
--save_precision "fp16"
--save_every_n_epochs 2
--cache_latents to_disk
--xformers
--lowram
</code></pre><pre><code>[训练参数]
BASE=https://civarchive.com/models/827184/wai-nsfw-illustrious-sdxl =&gt; v12
EPOCHS=8 # 训练轮数
SEED=1861

[分辨率设置]
TRAINING_WIDTH=1024 # 训练宽度
TRAINING_HEIGHT=1152 # 训练高度
MAX_BUCKET_RESO=$TRAINING_HEIGHT # 最大分桶分辨率

[硬件适配]
DIM=32 # 网络维度 
DIM_ALPHA=28 # 网络alpha值 
TRAINING_BATCH_SIZE=1 # 批次大小

[训练参数]
--train_batch_size 1
--learning_rate 0.0001
--unet_lr 0.0001
--text_encoder_lr 0.00001
--optimizer_type "AdamW8bit"
--mixed_precision "bf16"
--save_model_as "safetensors"
--save_precision "fp16"
--save_every_n_epochs 2
--cache_latents to_disk
--xformers
--lowram</code></pre>

Lab @1 : trying to have better pantyhose, see-through, maturity.

lora_lab_p24_wor.safetensors

v1.0

<h2 id="main-thread-of-this-series-is-e4g6d0hj2">Main Thread of this series is</h2><h3 id="https:civitai.commodels1362442-def0g2509">👉👉👉 <a target="_blank" rel="ugc" href="https://civarchive.com/models/1362442">https://civarchive.com/models/1362442</a></h3>Please visit the model above for more info. ( e.g. training params, random thoughts, ... )<h2 id="-v1omj51y7">这个实验系列的主题贴是</h2><h3 id="https:civitai.commodels1362442-t65owm9is">👉👉👉 <a target="_blank" rel="ugc" href="https://civarchive.com/models/1362442">https://civarchive.com/models/1362442</a></h3>请移步上述的 model 主页, 了解更多信息 (例如, 训练参数, 一些想法, ... )

Files with SHA256: dbeef2e68844be2ff15bc801f794ba86de184af574c708fb9b1a890fa56f3b56