A.I. vs A.I

Recently, the quality of AI-generated illustrations has improved dramatically, making it difficult to distinguish them from human-drawn works at a glance. This raised the question: 'Perhaps it is not humans, but AI itself, that can most accurately identify AI-drawn pictures?'

Therefore, I built an AI to distinguish between images generated by NovelAI and those drawn by humans. For training data, I used about 12,000 images each of 'NovelAI Girl' and 'Original Girl' (with over 300 views) collected from Pixiv, fine-tuning ResNet152. Random resizing and horizontal flipping were added as data augmentation.

As a result, the classification accuracy reached about 97%, and it correctly answered 42 out of 45 questions in the popular 'Human or AI' quiz, significantly exceeding the human average. ViT (Vision Transformer) achieved similar accuracy but had a lower quiz score than ResNet.

Interestingly, there were cases where ResNet and ViT made opposite judgments on the same image. This might be because ViT focuses more on shape than CNN and is less sensitive to texture differences.