Wondering how it works..?
Hi, how does this work, the training etc
thanks, really awesome network. works 9.5/10 times accurately..!
The training dataset is danbooru. And use 2 aesthetic model to label the images.
model1: https://github.com/waifu-diffusion/aesthetic
model2: https://huggingface.co/skytnt/anime-aesthetic/tree/4659f8d521aaf1b6f89c57f4d182154e2445d3fb
The 2 models is binary classification model but not score prediction model. I think they may make mistake when an input image is mediocre.
So I combine the score of the 2 models.
score = 0.5score1 + 0.5score2
Use the combined score to train my aesthetic score prediction model.
Loss function is mse loss instead of bce loss.
Dataset of https://huggingface.co/skytnt/anime-aesthetic/tree/4659f8d521aaf1b6f89c57f4d182154e2445d3fb is pixiv, just binary classification of excellent illustrators and bad illustrators.