Dantaggen Beta

🚀 DanTagGen - beta

DanTagGen (Danbooru Tag Generator) is a text - generation project inspired by p1atdev's dart project. It offers a unique approach with a different architecture, dataset, format, and training strategy, aiming to enhance the tag - generation process for Danbooru - related content.

✨ Features

• Version Differences:
  • Alpha: Pretrained on a 2M dataset with a smaller batch size, resulting in limited generation ability.
  • Beta: Pretrained on a 5.3M dataset with a larger batch size, offering more stability and better performance even with minimal input information.
• Model Architecture: Trained from scratch using a 400M - parameter LLaMA architecture (referred to as NanoLLaMA). It is theoretically compatible with any LLaMA inference interface.
• Model Formats: The repository provides converted FP16 gguf models and quantized 8 - bit/6 - bit gguf models, recommended to be run with llama.cpp or llama - cpp - python for high - speed performance.
• Dataset and Training: Trained for 10 epochs on 5.3M data, having seen approximately 6 - 12B tokens. The dataset is exported from HakuBooru using a custom - filtered Danbooru sqlite database.
• Utilities: A Gradio UI is under development, and plans are in place to create an sd - webui extension for wider application.

📦 Installation

No installation steps are provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

Example 1: Vivlos

Base prompt:

1girl,
vivlos \(umamusume\), umamusume, 
kxl - delta - style1,
swimsuit,
masterpiece, newest, absurdres, sensitive

Example 2: Daring Tact

Base prompt:

1girl,
daring tact \(umamusume\), umamusume, 
kxl - delta - style1,
horse girl, horse tail, horse ears, cafe, table, chair,
masterpiece, newest, absurdres, safe

Advanced Usage

The model uses the following format for input prompts:

prompt = f"""
rating: {rating or '<|empty|>'}
artist: {artist.strip() or '<|empty|>'}
characters: {characters.strip() or '<|empty|>'}
copyrights: {copyrights.strip() or '<|empty|>'}
aspect ratio: {f"{aspect_ratio:.1f}" or '<|empty|>'}
target: {'<|' + target + '|>' if target else '<|long|>'}
general: {", ".join(special_tags)}, {general.strip().strip(",")}<|input_end|>
"""

For example:

rating: safe
artist: <|empty|>
characters: <|empty|>
copyrights: <|empty|>
aspect ratio: 1.0
target: <|short|>
general: 1girl, solo, dragon girl, dragon horns, dragon tail<|input_end|>

And you may get an output like:

rating: safe
artist: <|empty|>
characters: <|empty|>
copyrights: <|empty|>
aspect ratio: 1.0
target: <|short|>
general: 1girl, solo, dragon girl, dragon horns, dragon tail<|input_end|>open mouth, red eyes, long hair, pointy ears, tail, black hair, chinese clothes, simple background, dragon, hair between eyes, horns, china dress, dress, looking at viewer, breasts

📚 Documentation

Model Arch

This version of DTG is trained from scratch using a 400M - parameter LLaMA architecture (referred to as NanoLLaMA). Since it is based on the LLaMA architecture, it can theoretically be used in any LLaMA inference interface. The repository also provides converted FP16 gguf models and quantized 8 - bit/6 - bit gguf models. It is recommended to use llama.cpp or llama - cpp - python to run this model for high - speed performance.

Dataset and Training

The training is conducted using a custom - implemented trainer in HakuPhi. The model is trained for 10 epochs on 5.3M data, having seen approximately 6 - 12B tokens. The dataset is exported from HakuBooru using a custom - filtered Danbooru sqlite database. The data is filtered based on the percentile of the fav_count for each rating (2M = top 25%, 5.3M = top 75%).

Utilities

A Gradio UI is currently being implemented for this project. Other developers can utilize the API within it to create different applications. Additionally, there are plans to develop an sd - webui extension.

📄 License

⚠️ Important Note

This project is tagged with "not - for - all - audiences" and "art". Please use it in compliance with relevant regulations and ethical standards.

💡 Usage Tip

When using the model, try different input prompts to get the best tag - generation results. You can adjust the rating, artist, characters, etc., according to your needs.