Introduction

This LoRA is trained on Shinano (しなの, 信浓), a VRChat avatar created by Ponderogen (ぽんでろ).

Usage

Recommended LoRA strength: 0.8-1.0
Trigger word: shinano (vrchat)

In ComfyUI, escape the parentheses:

shinano \(vrchat\)

Basic appearance:

grey hair, blue eyes, cat ears, side braid, black cross hairclips, cat tail

Default outfit:

black sleeveless high-neck blouse, black ribbon bow at the neck, grey pleated skirt, black pantyhose, brown ankle boots, white off-shoulder cable-knit cardigan, wide folded white ribbed collar, vertical twisted cable-knit panels, ribbed cuffs and hem, gold buttons, white front belt tab with a rectangular gold buckle, two long flat tan straps hanging below the skirt

Default outfit (without cardigan):

black sleeveless high-neck blouse, black ribbon bow at the neck, grey pleated skirt, grey belt, black pantyhose, brown ankle boots, bare shoulders

Training Details

• Training tool: sorryhyun/anima_lora v1.17.1

• Dataset: 11 manually captioned files, including 7 selected source images and 4 detail crops

• LoRA rank / alpha: 64 / 256

• Learning rate: 2e-5

• Optimizer: AdamW

• Learning rate schedule: Cosine decay with 5% warmup

• Training duration: 900 steps

• Precision: BF16

• Resolution: Aspect-ratio buckets, approximately 1 megapixel

• GPU: NVIDIA GeForce RTX 3090

Note: v0.9 is a beta version. To recover her appearance and settings precisely, it still carries 3D modeling vibes. I'm collecting more illustration images and hopefully it would get better in the next version.

简介

这个 LoRA 基于 Shinano（しなの、信浓）进行训练。Shinano 是由 Ponderogen（ぽんでろ）制作的 VRChat 虚拟形象。

使用

推荐 LoRA 权重： 0.8–1.0
激活词： shinano (vrchat)

在 ComfyUI 中，请对括号进行转义：

shinano \(vrchat\)

基础外观：

grey hair, blue eyes, cat ears, side braid, black cross hairclips, cat tail

默认服装：

black sleeveless high-neck blouse, black ribbon bow at the neck, grey pleated skirt, black pantyhose, brown ankle boots, white off-shoulder cable-knit cardigan, wide folded white ribbed collar, vertical twisted cable-knit panels, ribbed cuffs and hem, gold buttons, white front belt tab with a rectangular gold buckle, two long flat tan straps hanging below the skirt

默认服装（不穿外套）：

black sleeveless high-neck blouse, black ribbon bow at the neck, grey pleated skirt, grey belt, black pantyhose, brown ankle boots, bare shoulders

训练参数

• 训练工具： sorryhyun/anima_lora v1.17.1

• 训练数据： 11 个手工标注文件，包括 7 张精选源图和 4 张服装细节裁图

• LoRA rank / alpha： 64 / 256

• 学习率： 2e-5

• 优化器： AdamW

• 学习率调度： 余弦衰减，预热比例为 5%

• 训练量： 900 steps

• 训练精度： BF16

• 训练分辨率： 按长宽比分桶，约 1 百万像素

• 训练显卡： NVIDIA GeForce RTX 3090

注意: v0.9 是一个测试版本。为了精确还原她的外观，仍然有比较强的 3D 味。我正在收集更多的平面插画图片，希望能在下一个版本变得更好。