README.md

metadata

language:
  - en
pipeline_tag: image-classification
tags:
  - pytorch
  - clip
  - generated-image-detection
  - image-forensics

🔍 DoU — Diversity over Uniformity

Diversity over Uniformity: Rethinking Representation in Generated Image Detection
CVPR 2026

📄 Paper · 💻 GitHub · 🤗 Weights

DoU detects AI-generated images by preserving diverse and complementary forensic cues, reducing reliance on a few dominant features.

🧠 Model Overview

Built on CLIP ViT-L/14, DoU combines:

• Cue Information Bottleneck (CIB): filters task-irrelevant information from multi-layer visual features.
• Adaptive Feature Correlation Learning (AFCL): reduces redundancy among forensic cues during training.
• Learnable aggregation and prompts: combine complementary cues for real/generated classification.

Input	Output
RGB image, preprocessed to 224 × 224	0: Real · 1: Generated

The generated-image score is softmax(logits)[:, 1].

🚀 Quick Start

1. Download and install

git clone https://huggingface.co/Yanmou/DoU
cd DoU
python -m pip install -r requirements.txt

Install a compatible PyTorch/torchvision pair for your platform. Model initialization also requires the original OpenAI CLIP backbone, downloaded automatically on first use.

2. Run inference

python infer.py --repo-id Yanmou/DoU --image example.jpg --device cpu --seed 42

For GPU inference, use --device cuda.

Using a local checkpoint:

python infer.py --ckpt best.pth --image example.jpg --device cpu --seed 42

Note: This is a custom PyTorch model; use the provided inference script rather than a Transformers pipeline.

🖼️ Preprocessing

The evaluation transform follows the official implementation:

1. Convert to RGB.
2. If either dimension is below 224, tile the image until both dimensions reach at least 224.
3. Center-crop to 224 × 224, convert to a tensor, and apply CLIP normalization.

mean = [0.48145466, 0.4578275, 0.40821073]
std  = [0.26862954, 0.26130258, 0.27577711]

No resizing is applied. A standard CLIP resize-and-crop processor is therefore not equivalent.

📊 Benchmark Evaluation

Organize the evaluation data as follows:

test_root/
├── progan/
│   ├── 0_real/
│   └── 1_fake/
└── stylegan/
    ├── 0_real/
    └── 1_fake/

From the upstream/ directory, run:

python test.py --ckpt ../best.pth --data-root /path/to/test_root --datasets progan stylegan --device cpu --batch-size 16 --num-workers 0 --output result.xlsx

The script reports ACC, Real ACC, Fake ACC, F1, and AP, and saves results to Excel.

⚙️ Implementation Notes

• Checkpoint: a PyTorch dictionary containing the state dictionary under ckpt["model"].
• Inference: the variational bottleneck samples features even in eval() mode. The inference script sets a seed; outputs may still vary across devices and environments.
• Source snapshot: upstream/ contains official code from commit 230856fabd137e00580d5457b3b1ce7b664bcf99.
• Scope: binary image detection for research. Scores are not calibrated authenticity probabilities, and predictions do not establish image provenance.

📜 License

See upstream/LICENSE for the source code license and WEIGHTS_LICENSE.txt for the checkpoint terms.

📚 Citation

If you find this work useful, please cite:

@inproceedings{he2026diversity,
  title={Diversity over uniformity: Rethinking representation in generated image detection},
  author={He, Qinghui and Zhang, Haifeng and Qin, Qiao and Liu, Bo and Bi, Xiuli and Xiao, Bin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={40407--40417},
  year={2026}
}