AI Anime Image Detector ViT

🚀 AI Anime Image Detector ViT

This is a proof - of - concept model designed to detect anime - style AI images. By leveraging the Vision Transformer architecture, it offers a reliable solution for distinguishing between human - made and AI - generated anime images.

🚀 Quick Start

The model was trained on a dataset consisting of 1M human - made real anime images and 217K AI - generated anime images. To prevent biases, an equal number of each type of image was used during training. It was trained on a single RTX 3090 GPU for approximately 40 hours, going through about 35 epochs. You can access the training logs on my wandb.

✨ Features

• Anime Image Detection: Specialized in detecting anime - style AI images.
• Vision Transformer Architecture: Utilizes the power of Vision Transformer for accurate classification.
• Balanced Training: Trained with an equal amount of human - made and AI - generated images to avoid biases.

📚 Documentation

Evaluation

Each checkpoint of the model was evaluated on a dataset of 500 real and 500 AI - generated images. The final results are as follows:

• Training Loss: 0.1009
• Eval Loss: 0.1386

Using random crops seemed to enhance the model's generalization ability. However, since the training dataset only contained 512x512 images, bilinear interpolation was applied to every cropped image. Training the model on 1024x1024 images might further improve its performance.

Performance comparison

We conducted a small evaluation test on approximately 5000 images using currently available AI image detectors. Note that these models were not specifically trained on anime images.

💻 Usage Examples

Basic Usage

from transformers import AutoModelForImageClassification, AutoFeatureExtractor
import torch
from PIL import Image

model = AutoModelForImageClassification.from_pretrained("legekka/AI-Anime-Image-Detector-ViT")
feature_extractor = AutoFeatureExtractor.from_pretrained("legekka/AI-Anime-Image-Detector-ViT")

model.eval()

image = Image.open("example.jpg")
inputs = feature_extractor(images=image, return_tensors="pt")

outputs = model(**inputs)
logits = outputs.logits

label = model.config.id2label[torch.argmax(logits).item()]
confidence = torch.nn.functional.softmax(logits, dim=1)[0][torch.argmax(logits)].item()

print(f"Prediction: {label} ({round(confidence * 100)}%)")

📄 License

This project is licensed under the Apache 2.0 license.