Introduction

This is a concept/style LoRA for generating animegao kigurumi photography: performers wearing full-head anime character masks and costumes. It focuses on the distinctive combination of a rigid character head, a fixed anime-style expression, a wig, and a human-proportioned costumed body.

Anima is not particularly strong at rendering photorealistic images, so the results may retain a noticeable cartoon-like appearance. Combining this with a realism-enhancing LoRA might help, but I have not tested that yet.

Usage

Recommended LoRA strength: 1.0

Trigger word:

animekig

Describe the character, outfit, pose, and setting after the trigger word. For example:

animekig, photorealistic, solo, pink wig, blue eyes, maid dress, white thighhighs, standing in a studio

animekig, photorealistic, two performers, colorful idol costumes, posing together on a stage

Training Details

• Base model: Anima Base 1.0

• Training tool: sorryhyun/anima_lora v1.17.1

• Dataset: 162 curated and captioned photographs

• Captioning model: JoyCaption Beta One Q8

• LoRA rank / alpha: 32 / 128

• Learning rate: 2e-5

• Optimizer: AdamW

• Learning rate schedule: Cosine decay with 5% warmup

• Training duration: 8 epochs / 1,296 steps

• Released checkpoint: Epoch 8

• Precision: BF16

• Resolution: Aspect-ratio buckets, approximately 1 megapixel

• GPU: NVIDIA GeForce RTX 3090

简介

这是一个用于生成 Animegao Kigurumi 摄影图像的风格/概念 LoRA，也就是由真人穿戴动漫角色头壳及角色服装进行扮演的形式。它主要学习硬质角色头壳、固定的动漫面部表情、假发，以及真人比例服装身体之间的视觉组合。

Anima 对真人图片的表现并不特别好，因此生成结果仍可能带有比较明显的卡通感和涂抹感。搭配一些强化写实效果的 LoRA 也许会有所改善，不过我还没有实际测试过。

使用

推荐权重：1.0

激活词：

animekig

在激活词后描述角色外观、服装、姿势和场景。例如：

animekig, photorealistic, solo, pink wig, blue eyes, maid dress, white thighhighs, standing in a studio

animekig, photorealistic, two performers, colorful idol costumes, posing together on a stage

训练参数

• 基础模型：Anima Base 1.0

• 训练工具：sorryhyun/anima_lora v1.17.1

• 训练数据：162 张经过筛选和标注的照片

• 标注模型：JoyCaption Beta One Q8

• LoRA rank / alpha：32 / 128

• 学习率：2e-5

• 优化器：AdamW

• 学习率调度：余弦衰减，预热比例 5%

• 训练量：8 epochs / 1,296 steps

• 发布版本：第 8 epoch

• 训练精度：BF16

• 训练分辨率：按长宽比分桶，约 1 M 像素

• 训练显卡：NVIDIA GeForce RTX 3090