PixAI released a model for tagging anime images

Company PixAI introduced pixai-tagger-v1.0 — a specialized model designed for automatic recognition and classification of tags on anime images. The release is focused on improving metadata accuracy in visual generative projects.

The model's architecture contains 486 million parameters. The system is trained to recognize 30,877 unique tags distributed across 6 categories. This allows covering a wide range of visual characteristics, from stylistic features to specific character attributes.

In internal tests, the model demonstrated higher F1-score metrics compared to the previous version v0.9. Optimization was performed with images at a resolution of 1008x1008 pixels, providing high detail when analyzing complex compositions.

The update is aimed at improving the quality of data labeling for training diffusion models. Using more accurate tags allows reducing the number of errors when generating images from text prompts and increasing the result's compliance with the specified parameters.