WD 1.4 MOAT Tagger V2
| Property | Value |
|---|---|
| Framework | TF-Keras, ONNX |
| License | Apache-2.0 |
| Paper | MOAT: Alternating Mobile Convolution and Attention |
| Validation F1 Score | 0.6911 |
What is wd-v1-4-moat-tagger-v2?
The WD 1.4 MOAT Tagger V2 is an advanced image tagging model that leverages the MOAT (Mobile Convolution and Attention) architecture to provide comprehensive image tagging capabilities. Developed by SmilingWolf, this model has been specifically trained on the Danbooru dataset to recognize and tag images with ratings, characters, and general attributes.
Implementation Details
The model was trained using a carefully curated subset of the Danbooru dataset, specifically images with IDs modulo 0000-0899, while validation was performed on images with IDs modulo 0950-0999. The training process incorporated several key filtering criteria to ensure quality:
- Images with fewer than 10 general tags were excluded
- Tags appearing in fewer than 600 images were filtered out
- Training utilized TPUs provided by the TRC program
- Achieved optimal performance with a threshold of 0.3771
Core Capabilities
- Multi-category tagging support (ratings, characters, general tags)
- High-precision tag recognition with F1 score of 0.6911
- Efficient processing through TF-Keras and ONNX compatibility
- Robust performance on diverse image content
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its implementation of the MOAT architecture, which combines mobile convolution with attention mechanisms to achieve strong vision modeling capabilities. Its training on a carefully curated Danbooru dataset and high F1 score make it particularly effective for anime and illustration tagging tasks.
Q: What are the recommended use cases?
The model is ideal for automated image tagging systems, particularly those dealing with anime-style artwork and illustrations. It's especially useful for content management systems, digital art platforms, and image organization tools that require accurate tag prediction.
