v6 Update

数据集加入近期图片，重新打tag+自然语言打标训练，尤其是标注了年份（year 2021~2026），修改年份tag可能改变画风，但还是主要靠正面负面的质量提示词，质量越“好”越偏向厚涂，少写就偏向3d cg。

注意下载列表有多个变体，选择你心仪的epoch！

Add recent images to the dataset, re-tag and train with natural language labeling, especially with year annotations (year 2021–2026). Modifying the year tag may change the art style, but the primary control still relies on positive and negative quality prompts. Higher "quality" tends to lean toward a thick-paint (painterly) style, while less detailed prompts lean toward a 3D CG look.

Please note that there are multiple variants in the download list—choose the epoch you prefer!

History

v5

触发词@appa5ar7，这个lora着重训练了Appasart的伪3d画作，训练集内厚涂风格较少。也可以加@appas（Anima其实认得），不过不保证能增加质感，并且会增加出现水印的几率。

下载列表里有分别从50到75 epochs的lora，自行取用，因为我测不出来哪个比较好😭

测试下来这个lora对于2人场景下的动作经常会过拟合成插入性爱

Trigger word: @appa5ar7. This LoRa model was primarily trained on Appasart’s pseudo-3D artwork, and there are few thick-painting styles in the training dataset. You can also add @appas (Anima actually recognizes it), but there’s no guarantee it will improve the texture, and it will increase the likelihood of watermarks appearing.

The download list includes LoRa models ranging from 50 to 75 epochs—feel free to use whichever one you like, since I couldn’t figure out which one works best 😭

Based on the test, this LoRa often over-interprets movements in two-person scenarios as penetrative sex.

Please note that there are multiple variants in the download list—choose the epoch you prefer!