2023/2/15 update

Update dataset and model. Use offset noise to improve ability to generate lighter/darker images. Reduce the model file size to 1MB.

2023/2/8 update

Decrease the learning rate of text encoder to prevent overfitting.

Introduction

A LORA model of Taihou (azur lane) trained with 360 highres images, including nsfw and alternate costume, such as enraptured companion. The NAI's VAE is recommended. 0.6~0.8 weights are good.