camie-tagger

What is camie-tagger?

camie-tagger is an advanced deep learning model designed specifically for automatically tagging anime and manga illustrations. Built using a novel two-stage prediction approach, it can identify and classify content across 70,527 possible tags spanning seven distinct categories. The model achieves an impressive 61% F1 score, making it particularly effective for automated content analysis of anime-style artwork.

Implementation Details

The model employs a sophisticated architecture combining an EfficientNet V2-L backbone with a unique two-stage prediction system. The initial stage performs direct classification through a multi-layer classifier, while the refined stage uses cross-attention mechanisms to improve prediction accuracy. Notable is its ability to train effectively on consumer hardware (single RTX 3060) through optimized DeepSpeed configurations and specialized UnifiedFocalLoss function.

• Dual-mode operation supporting both full and initial-only predictions
• Comprehensive tag coverage across 7 categories including general, character, copyright, artist, meta, rating, and year
• Efficient training implementation requiring only 12GB VRAM
• Specialized loss function handling extreme class imbalance

Core Capabilities

• Character Recognition: 74.7% F1 score across 26,968 character tags
• Copyright Detection: 78.5% F1 score for identifying source materials
• General Tag Classification: 60.8% F1 score across 30,841 general tags
• Multiple threshold profiles for precision-recall tradeoff optimization
• Windows compatibility through initial-only mode