Skip to content

Latest commit

 

History

History
58 lines (40 loc) · 2.92 KB

README_EN.md

File metadata and controls

58 lines (40 loc) · 2.92 KB

nai3_train

If you need to use the images generated by nai3 to train your SD model in batches, then I believe you need this project.

The nai3-generated images are generated, labeled, and processed with a single script.

Principle: Randomly select a .txt file from the prompt_folder as a prompt to generate random materials in batches.

Note: If there is no pressure to update, there won't be any updates. The clear.py script has some issues, it can only remove keywords but not commas.

Open Source Random Prompt Word Library

A massive collection of random prompt words in .txt format. After extracting, place them in the prompt_folder.

  1. 4K High-Quality Prompt Word Library: Download

    • Features: Manually curated, high-quality, high safety rating with only 2% NSFW content.
    • Content: Mainly focused on characters and composition.
    • Source: Modified from the labeled files of the AID training set by Yugi.

  2. 200K High-Quality Prompt Word Library: Download

    • Features: Reverse-engineered from real images, large quantity, high concentration of 2D content, sufficient generalization, approximately 20% NSFW content.
    • Content: Diverse themes, considering the composition of Danbooru, predominantly female-focused.
    • Source: Batch crawled from https://danbooru.donmai.us/ and modified from the 200K training set by GangGangGe.

1. Batch Image Generation - Generating Random Images

Required Parameters:

  • self.token: Token required for image generation, authorization token. Obtain it as follows:

    • Log in to your NovelAI account on the website (https://novelai.net).
    • Open the console (F12) and switch to the Console tab.
    • Enter the following code and press Enter: 
      console.log(JSON.parse(localStorage.session).auth_token)
    • The output string is your authorization token.

  • prompt_folder: Randomly extract TXT files from this folder as prompts.

  • prefix: Default prefix, customized quality words or fixed artist styles.

  • negative_prompt: Negative prompt words.

  • num_images: Total number of generated images.

  • batch_size: Number of images generated per batch.

  • sleep_time: Sleep time after each batch generation (in seconds).

  • retry_delay: Time for the script to automatically restart after encountering an exception (in seconds).

Command to run: Python random_prompt

2. Read Image Information and Generate TXT Files with the Same Names

Command to run: Python nai3tagger.py

Then enter the folder path you want to process as prompted.

3. Remove the Prefix prefix from the TXT Files and Bind the Art Style

Command to run: Python clear.py

4. Remove Illegal Characters from the TXT Files, such as 😄🙃

Command to run: Python UTF-8.py