Skip to content

Latest commit

 

History

History
133 lines (82 loc) · 4.93 KB

README_EN.md

File metadata and controls

133 lines (82 loc) · 4.93 KB
Raw

✨ nai3_train ✨

English Version: English Documentation

If you need to use images generated by nai3 for bulk training of your SD model, this project is for you.

nai3 generates images, and this project provides a one-click script for generating, tagging, and processing these images.

Principle 1: Randomly extract text from the prompt_folder directory as a prompt to generate random materials in bulk.

Turn your tagged txt into your random prompt library!

Principle 2: Randomly extract characters from json to generate random characters, or generate all characters in order.

Note: New features are welcome, and the script is continuously being optimized.

💿 Installation and Running Environment

pip install pdm
# Automatically create a virtual environment and lock dependency versions
pdm install

⚙️ Configuration

Use a text editor such as Notepad to open the config.toml file in this directory, and configure it according to the comments. The token is required, and the rest can be left as default.

How to get the token:

  • Log in to your NovelAI account on the webpage (https://novelai.net)
  • Find "Account" in the top left personal center
  • Find the "Get Persistent API Token." section
  • Click the "Generate Token" button to generate a token

✨ Supported Samplers

"k_euler",
"k_euler_ancestral",
"k_dpmpp_2s_ancestral",
"k_dpmpp_2m",
"k_dpmpp_sde",
"ddim_v3",

✨ Generated Random Character Effects:

Image 1

Image1

Image 2

Image2

Image 3:

Image3

Image 4:

Image4

✨ Generated Random Material Effects:

Image 4:

Image4

Image 5:

Image5

🤗 Open Source Random Prompt Library

A massive txt random prompt library, just unzip and put it into prompt_folder.

Library 4K High Quality 200K High Quality
Features Manually sorted, high quality rate, high safety, only 2% nsfw Real image reverse push, large quantity, high concentration of secondary elements, enough generalization, about 20% nsfw
Content Mainly characters and composition Many themes, considering the composition of danbooru, mainly female characters
Source Modified from Yoshiji's AID training set tagging file Bulk crawled from https://danbooru.donmai.us/, modified from Brother Ganggang's 200K training set
Download 4K High Quality 200K High Quality

🎉 Quick Start

1️⃣ random_characetrs.py

Generate specified characters' images randomly or in order.

Principle: Read character names from json and generate images.

This repository includes random character libraries of several mainstream games including Genshin Impact, Arknights, FGO, etc.

Example: Use by reading the .\json\genshin.json file to generate random Genshin Impact characters.

The structure of the json file, taking genshin.json as an example:

{
  "role": [
    "noelle_(genshin_impact)",
    "faruzan_(genshin_impact)",
    "Character 1",
    "Character 2",
    "Character 3",
    .....
  ]

2️⃣ random_prompt.py

Read the random prompt library and generate random nai3 images in bulk.

3️⃣ nai3tagger.py

Read image information and generate a txt file with the same name.

4️⃣ clear.py

Remove the prefix prefix from the txt file to bind the style.

5️⃣ UTF-8.py

Remove illegal characters from the txt file, such as 😄🙃.

Some illegal characters in the txt file may cause errors during training, and the training script cannot read the txt, so removing illegal characters is necessary.

6️⃣ cut.py

Remove the watermark from the image (test feature).