This repository provides ONNX-optimized implementations of the ML-Danbooru image tagging models, originally developed by 7eu7d7. ML-Danbooru is a sophisticated deep learning system specifically designed for automated tagging of anime-style images, leveraging modern transformer architectures to achieve high-precision classification across thousands of Danbooru-style tags. The models in this repository have been converted to ONNX format for improved inference performance and cross-platform compatibility.
The core architecture employs Caformer (Convolution-Augmented Transformer) models, which combine the global receptive field of transformers with the local feature extraction capabilities of convolutional networks. This hybrid approach enables the models to effectively capture both fine-grained details and global contextual information in anime artwork. The repository includes multiple model variants trained with different configurations and epochs, providing users with options ranging from faster inference to higher accuracy depending on their specific requirements.
Performance-wise, these models demonstrate exceptional accuracy in recognizing common anime character attributes, clothing items, accessories, backgrounds, and compositional elements. They can reliably identify tags such as hair colors, eye colors, clothing types, character poses, and scene settings with confidence scores typically exceeding 0.7-0.9 for relevant features. The models support batch processing and can handle images of various aspect ratios through intelligent resizing strategies that preserve important visual information while maintaining computational efficiency.
Usage
The models in this repository are designed to be used with the dghs-imgutils library, which provides a comprehensive interface for image tagging tasks.
Installation
pip install dghs-imgutils
Basic Usage
python
1from imgutils.tagging import get_mldanbooru_tags
23# Tag an image with default settings4tags = get_mldanbooru_tags('your_image.jpg')5print(tags)67# Tag with custom threshold and settings8tags_custom = get_mldanbooru_tags(9'your_image.jpg',10 threshold=0.5,11 size=448,12 keep_ratio=True,13 drop_overlap=True,14 use_real_name=False15)16print(tags_custom)
Model Variants
This repository contains multiple ML-Danbooru model variants:
ml_caformer_m36_dec-5-97527.onnx: Primary model with Caformer-M36 architecture
ml_caformer_m36_dec-3-80000.onnx: Alternative checkpoint with different training
TResnet-D-FLq_ema_6-30000.onnx: Extended training TResnet variant
caformer_m36-3-80000.onnx: Base Caformer model
Tag Information
The repository includes comprehensive tag information:
classes.json: Contains 1,527 simplified tag names for common anime attributes
tags.csv: Complete tag database with 12,547 entries including:
Original tag names
Root forms for morphological variations
Part-of-speech classifications
Usage frequency counts
Performance Characteristics
Input Size: Default 448x448 pixels (configurable)
Tag Count: 12,547 possible tags
Threshold: Default 0.7 (configurable)
Supported Tags: Character attributes, clothing, accessories, backgrounds, compositions
Architecture: Caformer-M36 and TResnet variants
Format: ONNX for optimized inference
Model Architecture Details
The ML-Danbooru models utilize modern transformer-based architectures:
Caformer-M36: Combines convolutional layers with transformer blocks for efficient feature extraction
TResnet-D: Transformer-enhanced ResNet variants with focal loss optimization
ONNX Optimization: Models are exported with optimized operators for fast inference across different hardware platforms
Citation
bibtex
1@misc{deepghs_ml_danbooru_onnx,
2 title = {{ML-Danbooru ONNX Models: Optimized Anime Image Tagging}},
3 author = {7eu7d7 and DeepGHS Contributors},
4 howpublished = {\url{https://huggingface.co/deepghs/ml-danbooru-onnx}},
5 year = {2023},
6 note = {ONNX-optimized implementations of ML-Danbooru models for efficient anime image tagging with transformer-based architectures},
7 abstract = {This repository provides ONNX-optimized implementations of the ML-Danbooru image tagging models, originally developed by 7eu7d7. ML-Danbooru is a sophisticated deep learning system specifically designed for automated tagging of anime-style images, leveraging modern transformer architectures to achieve high-precision classification across thousands of Danbooru-style tags. The models employ Caformer (Convolution-Augmented Transformer) architectures that combine the global receptive field of transformers with local feature extraction capabilities of convolutional networks, enabling effective capture of both fine-grained details and global contextual information in anime artwork.},
8 keywords = {image-classification, anime, tagging, danbooru, transformer, onnx}
9}
