CunnyDetect is a state-of-the-art image classifier running entirely in the browser that aims to solve the age-old conundrum: "Is this thing a loli?"
- Can detect the presence of loli imagery with sufficient accuracy if the weather is good. (Lower accuracy if it's raining outside or if you look at it funny.)
- Lightweight. Can run in the browser.
- Help convince your peers that your waifu is, in fact, not a loli.
prediction: cunny 😭
prediction: not_cunny 😇
| Model | Number of images | Validation accuracy |
|---|---|---|
| mobilenetv3_finetuned(3M) | 26446 | 86.7 |
TensorFlow 2.10.0 is required if you want GPU support on Windows. Also cudatoolkit, cuda-nvcc, and cudnn.
If not training, CPU should be enough for prediction.
Training
Place images in training_data
python model.py
Predicting
Place images in images_to_predict
python predict.py
Converting to tensorflowjs model
tensorflowjs_converter --input_format=keras --output_format=tfjs_graph_model models/mobilenetv3_finetuned(3M).h5 client/js_model
Images are sourced from gelbooru, of which many are too cursed to redistribute.
| Class | Example tags |
|---|---|
| cunny | 1girl solo loli flat-chest female_focus highres |
| not_cunny | 1girl 1boy solo breasts mature_female female_focus male_focus highres |