noaione/pixai-tagger-v1.0-onnx
PixAI Tagger (ONNX)
This is an ONNX version of pixai-labs/pixai-tagger-v1.0. Model information, training details, intended use, and licensing follow the upstream model. The only difference is that this version is provided for ONNX Runtime inference.
This model is mainly built for noaione/pixtaggers-bot.
Tags
The model returns 30,877 logits in the same order as the upstream tagger. tags.json contains the tag names grouped into the same categories:
general, character, copyright, style, meta, rating
Each category stores:
- offset: the first global model index for that category.
- count: the number of tags in the category.
- tags: the ordered tag names for that category.
To resolve a model output index:
global index = category offset + local tag position
The runner applies sigmoid to the logits and returns the resolved tag name, global index, raw logit, and probability. Default category thresholds are kept in run_onnx.py.
Example Usage
Install the runtime dependencies:
pip install numpy onnxruntime Pillow
Run the tagger:
python run_onnx.py image.jpg
The ONNX model and its accompanying model.onnx.data file must remain together.
License
Same as the base model, this model is licensed under the Apache License 2.0.