[VTuber] Amakami Konomi | 甘狼このみ | MiliPro | ミリプロ

Introduction

This is my first LoRA. Amakami Konomi is a Japanese female indie VTuber, illustrator, and Live2D modeler. I find her illustrations really nice, so I decided to train this.

Usage

Recommended LoRA strength: 1.0

Trigger word:

amakami konomi

Default outfit:

blue dress, brown capelet, white frilled blouse, brown neck bow, belt, brown striped thighhighs, brown lace-up boots

Training Details

• Base model: Anima Base 1.0

• Training tool: sorryhyun/anima_lora v1.17.1

• Dataset: 7 manually captioned images

• LoRA rank / alpha: 32 / 128

• Learning rate: 2e-5

• Optimizer: AdamW

• Learning rate schedule: Cosine decay with 5% warmup

• Training duration: 120 epochs / 840 steps

• Precision: BF16

• Resolution: Aspect-ratio buckets, approximately 1 megapixel

• GPU: NVIDIA GeForce RTX 3090

简介

这是我的第一个 Lora。甘狼このみ 是一个日本的女性独立 VTuber/插画家/Live2D 建模师。我觉得她的插画很可爱，所以我训练了这个 Lora。

使用

激活权重：1.0

激活词：

amakami konomi

常服：

blue dress, brown capelet, white frilled blouse, brown neck bow, belt, brown striped thighhighs, brown lace-up boots

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：7 张手工标注图片

• 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：120 epochs / 840 steps

• 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