AI-Generated Image Detection: A Technical Comparative Study of Neural Network and Physical Photometry Methods

Project Basic Information

• Original Author/Maintainer: DrStrangel0ve
• Source Platform: GitHub
• Release Date: June 12, 2026

Core Content

This open-source project systematically compares neural network classifiers and physical photometry-based methods for AI-generated image detection, explores the technical route of fusing deep learning with traditional physical features, and provides a reproducible research framework for the field of AI-generated image forensics.

With the development of generative AI technologies such as Stable Diffusion, DALL-E, and MidJourney, the quality of AI-generated images has become difficult to distinguish, posing social challenges like misinformation dissemination and identity fraud. Traditional detection methods fall into two categories:

1. Deep learning-based neural network classifiers (automatically learning statistical features)
2. Physical photometry-based analysis methods (verifying the consistency of physical attributes like lighting and shadows)

A single method is hard to handle diverse generative models and complex scenarios, so exploring a fusion route is necessary.

The project implements four types of detection methods:

1. Standard Neural Network Classifier: Uses ResNet-18 as the backbone, with end-to-end supervised learning to distinguish between real and generated images.
2. Photometry Normal Vector Consistency Detection: Judges authenticity through physical features such as local normal vector estimation and integrability test.
3. Traditional Digital Forensics Baseline: Includes classic methods like noise residual analysis and JPEG compression artifact detection.
4. Physics-Guided Neural Fusion Model: Fuses ResNet embeddings with features like physical photometry and noise residuals, and makes comprehensive judgments via MLP, balancing representational ability and interpretability.

Dataset

• Main Benchmark: Kaggle CIFAKE Dataset (real CIFAR-style images vs Stable Diffusion-generated images)
• Supports over 20 datasets: such as AI vs Real 2026, ChatGPT/Gemini Deepfake 2026, MS COCOAI 2026, etc.

Technical Implementation Details

• Single Image Photometry Proxy Methods: Local normal vector estimation, integrability test, high-frequency shadow artifact detection, etc.
• Fusion Model: Inputs ResNet-18 visual embeddings + combined_v3 forensics feature vectors, outputs judgments via MLP.

1. Fusion Outperforms Single Methods: Neural networks are prone to overfitting, while physical methods lack expressive power; fusion models improve cross-model generalization ability.
2. Value of Physical Constraints: In single-image scenarios, physical photometry constraints can capture systematic flaws in generated images.
3. Importance of Interpretability: In high-risk scenarios, physics-guided models provide decision-making basis and enhance credibility.
4. Necessity of Dataset Diversity: Cross-dataset validation is key to evaluating the practical value of methods.

Research Roadmap

The next-generation model SCP-Fusion will expand: retain the physical branch, introduce CLIP/DINO embeddings, extend combined_v4 features, and add reconstruction error features.

Publication Plan

• DFRWS-USA 2026 (Poster)
• IEEE WIFS 2026 (Full Paper)
• DFF-2026 at ACM Multimedia (Thematic Seminar)

Reproducibility Guarantee

Provide detailed checklists, environment configurations (CPU/GPU), requirements files, and CUDA installation guides.

This project provides a solid experimental platform and reproducible framework for the field of AI-generated image detection. The idea of fusing deep learning with physical laws is a promising direction for the field's development. The game between generation and detection will exist for a long time; the value of open-source projects lies in establishing transparent benchmarks and promoting the field toward a reliable, interpretable, and reproducible direction.