Introduction
If you can only see the cover image, you may need to enable at least R-rated content in your Civitai content settings. Or check out the comparison image here
This is a Visual Concept Slider LoRA for Anima that controls the apparent opacity and thickness of hosiery. It works with pantyhose, tights, stockings, thigh-highs, and similar garments. Negative weights make the hosiery progressively thinner and more transparent, while positive weights make it denser and more opaque.
✅ Works well with both SFW and NSFW subjects
✅ Supports different hosiery colors, such as red and yellow
✅ Plug and play; tested successfully on Anima Base and fine-tuned models such as WAI-Anima
⚠️ Preview version: Very high weights may slightly alter the composition (although no significant changes occurred in my testing) or shift the visual emphasis
Usage
Recommended weight: -1.0 to 1.0 (positive values make the hosiery thicker and more opaque; negative values make it thinner and more transparent)
This LoRA does not require a trigger word.
Training Details
Base model: Anima Base 1.0
Training tool:
sorryhyun/anima_lora v1.17.1(with a custom Paired Visual Concept Slider loss function!)Training dataset: 11 manually paired image sets
LoRA rank / alpha: 4 / 1
Learning rate:
2e-4Optimizer: AdamW
Learning rate schedule: Constant learning rate with no warmup
Training duration: Approximately 22 epochs / 240 paired updates
Locality preservation weight: 0.5
Released checkpoint: Step 240
Training precision: BF16
Training resolution: Aspect-ratio buckets, with a target resolution of approximately
1024 × 896GPU: NVIDIA GeForce RTX 3090
简介
如果只能看到标题图，你可能需要在 Civitai 内容设置中打开至少 R 等级。或者在这里查看对比图
这是一个适用于 Anima 的 Visual Concept Slider LoRA，用于控制袜类的视觉透明度和厚度表现。它适用于裤袜、连裤袜、丝袜和长筒袜等，可以让袜子在负权重下逐渐变得轻薄、透肤，在正权重下逐渐变得更加致密和不透明。
✅在 SFW 和 NSFW 主题下均工作良好
✅支持不同颜色的袜类（例如红色、黄色等）
✅即插即用，在 Anima Base 和 WAI-Anima 等微调上均测试有效
⚠️预览版：过高的权重可能会略微改变构图（在我的测试中，没有明显影响）或影响视觉重心
使用
推荐权重：-1.0 - 1.0（正权重为变厚/不透明，负权重为变薄/透明）
本 LoRA 没有任何激活词。
训练参数
基础模型：Anima Base 1.0
训练工具：
sorryhyun/anima_lora v1.17.1（自己编写了 Paired Visual Concept Slider Loss 函数！）训练数据：11 组人工配对图片
LoRA rank / alpha：4 / 1
学习率：
2e-4优化器：AdamW
学习率调度：恒定学习率，无预热
训练量：约 22 epochs / 240 paired updates
局部保持权重：0.5
发布版本：240-step
训练精度：BF16
训练分辨率：按长宽比分桶，目标分辨率约
1024 × 896训练显卡：NVIDIA GeForce RTX 3090
Description
FAQ
Comments (9)
This is great! Exactly what I was looking for. Do you have an Illustrious version as well?
Thanks.
"Do you have an Illustrious version as well?" - I'm considering training ones for other model series too!
I've published a preview version for IllustriousXL. Have a try. It feels less stable (comparing to Anima) but still works well.
@volda thank you! I'll try it out
This is great! It's so close to what I've been fighting forever. Can I ask how you trained this, the file size is incredible. Ive been trying forever yo get fully opaque socks but it always falls apart at the toes. I would like it to function like how actual thigh socks do where you can see the general shape but not each individual toe. Like in this image
Thanks! This is also something I've always wanted. (Adding "thick" to an Anima prompt usually just makes the thighs thicker... 😅)
Simply put, you need to use something like contrastive learning. The goal is to find a specific direction vector in weight space: one that points from the thinnest -> thickest hosiery features.
Because essentially, what you're trying to do here is reinforce some knowledge the model already has, the LoRA file can be very small (here we choose rank=4). You need to use a special loss function.
I might write an article explaining the whole process if I have time.
As for the toe part, I didn't specifically add close-up images of feet to the training set, so you're welcome to test it and give feedback!
@volda hey sorry to keep bothering you about this, so ive been trying to make the complimentary silder to this where it creates a solid toe box(no individual toes visible) and i was jsut wondering what kind of framing you used for your training images. if possible if i could see a pair or two of the matching sets it would be super helpful. thanks
@vazht1s No worries. I roughly understand the effect you're trying to achieve, but I'm not sure whether the workflow can be reused.
Most importantly, your image pairs should be almost pixel-identical (except for the feature you want to train). The way I did this was by using Gemini's Nano Banana to generate some images with modified features as pairs. I'd give it an illustration with thick thigh-highs, then ask it to "generate the corresponding 10D thin thigh-highs" or vice versa, something like that, and then regenerate until I got a satisfactory image.
Here are some of them, for example: https://drive.google.com/file/d/1PwFgnIs_ws14rzwHkAgu13VHdy87ole3/view
goooooooooooooooooood!!!