README.md · deepghs/ai_image

metadata

license: openrail
metrics:
  - accuracy
pipeline_tag: image-classification
tags:
  - art
datasets:
  - deepghs/ai_image_corrupted
library_name: dghs-imgutils

This is the classifier model for predicting the anime-style stable-diffusion-generated images are corrupted or not.

Trained on dataset deepghs/ai_image_corrupted.

Models

Name	FLOPS	Params	Accuracy	AUC	Confusion	Labels
caformer_s36_v0_focal	22.10G	37.21M	95.65%	0.9916	confusion	`corrupted`, `normal`
caformer_s36_v0_sce	22.10G	37.21M	95.77%	0.9894	confusion	`corrupted`, `normal`
mobilenetv3_v0_focal_dist	0.63G	4.18M	94.02%	0.9842	confusion	`corrupted`, `normal`

How to Use

You can use this model with dghs-imgutils.

from imgutils.generic import classify_predict_score

classify_predict_score(
    'sample_image.png',
    repo_id='deepghs/ai_image_corrupted',
    model_name='mobilenetv3_v0_focal_dist',
)
# {'corrupted': 0.7807788848876953, 'normal': 0.2192210853099823}

Note

This model is trained on SD1.5 images, generated by 8 different base models.

So the better solution for this problem is to use some metrics like artstyle embeddings, this model will be deprecated as soon as the artstyle embedding model for anime images is completed.

Citation

@misc{Citation,
  title={AI-Corrupt Score for Anime Images},
  author={narugo1992},
  year={2023},
  howpublished={\url{https://huggingface.co/deepghs/ai_image_corrupted}}
}