Camais03/camie-tagger-v2

Camie Tagger v2

An advanced deep learning model for automatically tagging anime/manga illustrations with relevant tags across multiple categories, achieving an amazing 67.3% micro F1 score (using the micro optimized threshold profile) and 50.6% macro F1 score (using the macro optimized threshold preset) across 70,527 possible tags on a test set of 20,116 samples. Now with Vision Transformer backbone and significantly improved performance. This dataset is notoriously long-tailed and sparse.

🚀 What's New in v2

Major Performance Improvements:

• Micro F1: 58.1% → 67.3% (+9.2 percentage points)
• Macro F1: 33.8% → 50.6% (+16.8 percentage points)
• Model Size: 424M → 143M parameters (-66% reduction)
• Architecture: Switched from EfficientNetV2-L to Vision Transformer (ViT) backbone
• Simplified Design: Streamlined from dual-stage to single refined prediction model

Training Innovations:

• Multi-Resolution Training: Progressive scaling from 384px → 512px resolution
• IRFS (Instance-Aware Repeat Factor Sampling): Significant macro F1 improvements for rare tags
• Adaptive Training: Models quickly adapt to resolution/distribution changes after initial pretraining
• Overall the model is more accurate, faster, and less code!

✨ Features:

• Streamlit web interface app and game: User-friendly UI for uploading and analyzing images and a tag collection game
• Adjustable threshold profiles: Micro, Macro, Balanced, Category-specific, profiles
• Fine-grained control: Per-category threshold adjustments for precision-recall tradeoffs
• Safetensors and ONNX: Available in main directory

📊 Performance Analysis:

Complete v1 vs v2 Performance Comparison:

CATEGORY	v1 Micro F1	v2 Micro F1	Micro Δ	v1 Macro F1	v2 Macro F1	Macro Δ
Overall	61.3%	67.3%	+6.0pp	33.8%	50.6%	+16.8pp
Artist	48.0%	70.0%	+22.0pp	29.9%	66.1%	+36.2pp
Character	75.7%	83.4%	+7.7pp	52.4%	66.2%	+13.8pp
Copyright	79.2%	86.6%	+7.4pp	41.9%	56.2%	+14.3pp
General	60.8%	66.4%	+5.6pp	21.5%	34.6%	+13.1pp
Meta	60.2%	61.2%	+1.0pp	14.5%	23.7%	+9.2pp
Rating	80.8%	83.1%	+2.3pp	79.5%	77.5%	-2.0pp
Year	33.2%	30.8%	-2.4pp	29.3%	32.6%	+3.3pp

Micro F1 comparison using micro-optimized thresholds, Macro F1 comparison using macro-optimized thresholds for fair evaluation.

Key Performance Insights:

The v2 model shows remarkable improvements across nearly all categories:

• Artist Recognition: Massive +22.0pp micro F1 improvement and +36.2pp macro improvement, indicating much better artist identification
• Character Detection: Large +7.7pp micro F1 and +13.8pp macro F1 gains
• Copyright Recognition: Excellent +7.4pp micro F1 improvement and +14.3pp macro improvement for series identification
• General Tags: Improved +5.6pp micro F1 and +13.1pp macro F1 for visual attributes
• Overall Macro F1: Exceptional +16.8pp improvement shows much better rare tag recognition

Only the year category shows slight regression.

Detailed v2 Performance:

MACRO OPTIMIZED (Recommended):

CATEGORY	THRESHOLD	MICRO-F1	MACRO-F1
overall	0.492	60.9%	50.6%
artist	0.492	62.3%	66.1%
character	0.492	79.9%	66.2%
copyright	0.492	81.8%	56.2%
general	0.492	60.2%	34.6%
meta	0.492	56.3%	23.7%
rating	0.492	78.7%	77.5%
year	0.492	37.2%	32.6%

MICRO OPTIMIZED:

CATEGORY	THRESHOLD	MICRO-F1	MACRO-F1
overall	0.614	67.3%	46.3%
artist	0.614	70.0%	64.4%
character	0.614	83.4%	64.5%
copyright	0.614	86.6%	53.1%
general	0.614	66.4%	27.4%
meta	0.614	61.2%	19.2%
rating	0.614	83.1%	81.8%
year	0.614	30.8%	21.3%

The model performs exceptionally well on character identification (83.4% F1 across 26,968 tags), copyright/series detection (86.6% F1 across 5,364 tags), and content rating classification (83.1% F1 across 4 tags).

Real-world Tag Accuracy:

The macro optimized threshold is recommended as many "false positives" according to the benchmark are actually correct tags missing from the Danbooru dataset. The model frequently identifies appropriate tags that weren't included in the original tagging, making perceived accuracy higher than formal metrics suggest.

If you'd like to support further training on the complete dataset or my future projects, consider supporting me here:(https://ko-fi.com/camais). Your support will directly enable longer training runs and better models!

🧠 Architecture Overview:

Vision Transformer Backbone:

• Base Model: Vision Transformer (ViT) with patch-based image processing
• Dual Output: Patch feature map + CLS token for comprehensive image understanding
• Efficient Design: 86.4M backbone parameters vs previous 214M+ classifier layers

Refined Prediction Pipeline:

1. Feature Extraction: ViT processes image into patch tokens and global CLS token
2. Global Pooling: Combines mean-pooled patches with CLS token (dual-pool approach)
3. Initial Predictions: Shared weights between tag embeddings and classification layer
4. Candidate Selection: Top-K tag selection based on initial confidence
5. Cross-Attention: Tag embeddings attend to image patch features
6. Final Scoring: Refined predictions for selected candidate tags

Key Improvements:

• Shared Weights: Tag embeddings directly used for initial classification
• Simplified Pipeline: Single refined prediction stage (vs previous initial + refined)
• Native PyTorch: Uses optimized MultiheadAttention instead of Flash Attention
• Custom Embeddings: No dependency on external models like CLIP
• Gradient Checkpointing: Memory-efficient training on consumer hardware

🛠️ Training Details:

Multi-Resolution Training Strategy:

The model was trained using a multi-resolution approach:

1. Phase 1: 3 epochs at 384px resolution with learning rate 1e-4
2. Phase 2: IRFS (Instance-Aware Repeat Factor Sampling) - addresses long-tailed distribution imbalance
3. Phase 3: 512px resolution fine-tuning with learning rate 5e-5

Key Training Insights:

Rapid Adaptation: Once the model learns good general features during initial pretraining, it adapts to resolution changes and distribution shifts very quickly - often within a fraction of an epoch rather than requiring full retraining.

IRFS Benefits: Instance-Aware Repeat Factor Sampling provided substantial macro F1 improvements by addressing the long-tailed distribution of anime tags, where instance counts vary dramatically between classes even with similar image counts.

Efficient Scaling: The ViT architecture generalizes resolution and capacity changes to the entire dataset, making incremental training highly efficient.

Training Data:

• Training subset: 2,000,000 images
• Training duration: 3+ epochs with multi-resolution scaling
• Final resolution: 512x512 pixels

🎮 Tag Collector Game (Camie Collector)

Introducing a Tagging game - a gamified approach to anime image tagging that helps you understand the performance and limits of the model. This was a shower thought gone to far! Lots of Project Moon references.

How to Play:

1. Upload an image
2. Scan for tags to discover them
3. Earn TagCoins for new discoveries
4. Spend TagCoins on upgrades to lower the threshold
5. Lower thresholds reveal rarer tags!
6. Collect sets of related tags for bonuses and reveal unique mosaics!
7. Visit the Library System to discover unique tags (not collect)
8. Use collected tags to either inspire new searches or generate essence
9. Use Enkephalin to generate Tag Essences
10. Use the Tag Essence Generator to collect the tag and related tags to it. Lamp Essence:

🖥️ Web Interface Guide

The interface is divided into three main sections:

1. Model Selection (Sidebar):

   • Choose between ONNX accelerated or safetensors model
   • View model information and memory usage

2. Image Upload (Left Panel):

   • Upload your own images or select from examples
   • View the selected image

3. Tagging Controls (Right Panel):

   • Select threshold profile
   • Adjust thresholds for precision-recall and micro/macro tradeoff
   • Configure display options
   • View predictions organized by category

Display Options:

• Show all tags: Display all tags including those below threshold
• Compact view: Hide progress bars for cleaner display
• Minimum confidence: Filter out low-confidence predictions
• Category selection: Choose which categories to include in the summary

Interface Screenshots:

Note the rare characters and tags idenified. Some only have 100's of samples on danbooru!

🛠️ Requirements

• Python 3.11.9 specifically (newer versions are incompatible)
• PyTorch 1.10+
• Streamlit
• PIL/Pillow
• NumPy

🔧 Usage

Setup the application and game by executing setup.bat. This installs the required virtual environment:

• Upload your own images or select from example images
• Choose different threshold profiles
• Adjust category-specific thresholds
• View predictions organized by category
• Filter and sort tags based on confidence

Use run_app.bat and run_game.bat.

🧠 Training Details

Dataset

The model was trained on a carefully filtered subset of the Danbooru 2024 dataset, which contains a vast collection of anime/manga illustrations with comprehensive tagging.

Filtering Process:

The dataset was filtered with the following constraints:

# Minimum tags per category required for each image
min_tag_counts = {
    'general': 25, 
    'character': 1, 
    'copyright': 1, 
    'artist': 0, 
    'meta': 0
}

# Minimum samples per tag required for tag to be included
min_tag_samples = {
    'general': 20, 
    'character': 40, 
    'copyright': 50, 
    'artist': 200, 
    'meta': 50
}

This filtering process:

1. First removed low-sample tags (tags with fewer occurrences than specified in min_tag_samples)
2. Then removed images with insufficient tags per category (as specified in min_tag_counts)

Training Data:

• Starting dataset size: ~3,000,000 filtered images
• Training subset: 2,000,000 images (due to storage and time constraints)

Preprocessing:

Images were preprocessed with minimal transformations:

• Tensor normalization (scaled to 0-1 range)
• ImageNet normalization.
• Resized while maintaining original aspect ratio
• No additional augmentations were applied

Tag Categories:

The model recognizes tags across these categories:

• General: Visual elements, concepts, clothing, etc. (30,841 tags)
• Character: Individual characters appearing in the image (26,968 tags)
• Copyright: Source material (anime, manga, game) (5,364 tags)
• Artist: Creator of the artwork (7,007 tags)
• Meta: Meta information about the image (323 tags)
• Rating: Content rating (4 tags)
• Year: Year of upload (20 tags)

All supported tags are stored in model/metadata.json, which maps tag IDs to their names and categories.

Training Notebooks

The repository includes the main training notebook:

1. camie-tagger-v2.ipynb:
   • Main training notebook
   • Dataset loading and preprocessing
   • Model initialization
   • Tag selection optimization
   • Metric tracking and visualization

Training Monitor

The project includes a real-time training monitor accessible via browser at localhost:5000 during training:

Performance Tips:

⚠️ Important: For optimal training speed, keep VSCode minimized and the training monitor open in your browser. This can improve iteration speed by 3-5x due to how the Windows/WSL graphics stack handles window focus and CUDA kernel execution.

Monitor Features:

The training monitor provides three main views:

1. Overview Tab:

• Training Progress: Real-time metrics including epoch, batch, speed, and time estimates
• Loss Chart: Training and validation loss visualization
• F1 Scores: Initial and refined F1 metrics for both training and validation

2. Predictions Tab:

• Image Preview: Shows the current sample being analyzed
• Prediction Controls: Toggle between initial and refined predictions
• Tag Analysis:
  • Color-coded tag results (correct, incorrect, missing)
  • Confidence visualization with probability bars
  • Category-based organization
  • Filtering options for error analysis

3. Selection Analysis Tab:

• Selection Metrics: Statistics on tag selection quality
  • Ground truth recall
  • Average probability for ground truth vs. non-ground truth tags
  • Unique tags selected
• Selection Graph: Trends in selection quality over time
• Selected Tags Details: Detailed view of model-selected tags with confidence scores

The monitor provides invaluable insights into how the two-stage prediction model is performing, particularly how the tag selection process is working between the initial and refined prediction stages.

Training Notes:

• Training notebooks may require WSL and 32GB+ of RAM to handle the dataset
• With more computational resources, the model could be trained longer on the full dataset

🙏 Acknowledgments

• Claude Sonnet 3.5, 4 and ChatGPT 5 Thinking for development assistance and brainstorming
• Vision Transformer for the foundational architecture
• Danbooru for the comprehensive tagged anime image dataset
• p1atdev for the processed Danbooru 2024 dataset
• IRFS paper for Instance-Aware Repeat Factor Sampling methodology
• PyTorch team for optimized attention implementations and gradient checkpointing
• The open-source ML community for foundational tools and methods