Modotte

• GitHub Repository
• Live Demo

Overview

In an era of rapidly advancing AI-generated imagery, deepfakes, and synthetic media, the need for reliable detection tools has never been higher. AIRealNet is a binary image classifier explicitly designed to distinguish AI-generated images from real human photographs. This model is optimized to detect conventional AI-generated content while adhering to strict privacy standards—avoiding personal or sensitive images.

• Class 0: AI-generated image
• Class 1: Real human image

By leveraging the robust SwinV2 Tiny architecture as its backbone, AIRealNet achieves a high degree of accuracy while remaining lightweight enough for practical deployment.

Key Features

1. High Accuracy on Public Datasets: Despite using a 14k-image fine-tuning split(Part of main fine tuning split), AIRealNet demonstrates exceptional accuracy and robustness in detecting AI-generated images.

2. Balanced Training Split: The dataset contains a balanced number of AI-generated and real images, ensuring unbiased training and minimizing class imbalance issues.

   • AI-Generated: 60%
   • Human-Images: 40%

3. Ethical Design: No personal photos were included, even if edited or AI-modified, respecting privacy and ethical AI principles.

4. Fast and Scalable: Based on a transformer vision model, AIRealNet can be deployed efficiently in both research and production environments.

Training Data

• Dataset: Parveshiiii/AI-vs-Real (open-sourced subset of main dataset )
• Size: 14k images (balanced between AI and human)
• Split: Used the train split for fine-tuning; validation performed on a separate balanced subset.
• Notes: Images sourced from public datasets and AI generation tools. Edited personal photos were intentionally excluded.

Limitations

While AIRealNet performs exceptionally well on typical AI-generated images, users should note:

1. Subtle Edits: The model struggles with nano-scale edits or ultra-precise modifications, like “nano banana” edits.
2. Edited Personal Images(over precise): Images of real people that have been AI-modified are not detected, aligning with privacy and ethical guidelines.
3. Domain Generalization: Performance may vary on images from completely unseen AI generators or extremely unconventional content.

Performance Metrics

Metrics shown are from Epoch 2, chosen to illustrate stable performance after fine-tuning.

Note: Extremely low loss and high accuracy are due to the controlled dataset environment. Real-world performance may be lower depending on the image domain.(In our testing this is model is over accurate despite it can't detect Nano-Banana images(only edited fully generated images can be detected over accurately))

Demo and Usage

1. Installing dependecies

pip install -U transformers

2. Loading and running a demo

from transformers import pipeline

pipe = pipeline("image-classification", model="Modotte/AIRealNet")
pipe("https://cdn-uploads.huggingface.co/production/uploads/677fcdf29b9a9863eba3f29f/eVkKUTdiInUl6pbIUghQC.png")# example image

Demo

• Given Image(Checkout Maths best filtered dataset focused on reasoning on Modotte)

• Model Output

[{'label': 'artificial', 'score': 0.9865425825119019},
 {'label': 'real', 'score': 0.013457471504807472}]

Note: its correct as the image was generated by a diffusion model

Intended Use

• Detect AI-generated imagery on social media, research publications, and digital media platforms.
• Assist content moderators, researchers, and fact-checkers in identifying synthetic media.
• Not intended for legal verification without human corroboration.

Ethical Considerations

• Privacy-first Approach: Personal photos, even if AI-edited, were excluded.
• Responsible Deployment: Users should combine model predictions with human review to avoid false positives or negatives.
• Transparency: The model card openly communicates its limitations and dataset design to prevent misuse.

How It Works

1. Images are preprocessed and resized to 256x256.
2. Features are extracted using the SwinV2 Tiny vision transformer backbone.
3. A binary classification head outputs probabilities for AI-generated vs real human images.
4. Predictions are interpreted as class 0 (AI) or class 1 (Human).

Future Work

Future iterations aim to:

• Improve detection of subtle AI-generated edits and “nano banana” modifications.
• Expand training data with diverse AI generators to enhance generalization.
• Explore multi-modal detection capabilities (e.g., video, metadata, and image combined).

Citation

@misc{Modotte_AIRealNet_2025,
    title={AIRealNet: A Fine-Tuned Vision Transformer for Detecting AI-Generated vs Real Human Images},
    author={Parvesh Rawal},
    publisher={Hugging Face},
    year={2025},
    url={https://huggingface.co/Modotte/AIRealNet}
}

References

• Microsoft SwinV2 Tiny: https://github.com/microsoft/Swin-Transformer
• Parveshiiii/AI-vs-Real dataset (subset): Open-sourced by our team member