🚀 DoU

Diversity over Uniformity: Rethinking Representation in Generated Image Detection

Official PyTorch implementation for generated image detection.

🧠 Overview

DoU revisits representation learning for generated image detection. Instead of enforcing representation uniformity, DoU promotes cue diversity while preserving robustness and generalization ability.

🔑 Core Components

Component	Full Name	Description
🧩 CIB	Cue Information Bottleneck	A VAE-style variational bottleneck applied to multi-stage visual cues.
🔄 AFCL	Adaptive Feature Correlation Learning	An inter-cue decorrelation mechanism that encourages diverse visual cues.
🖼️ Multi-Stage CLIP	Multi-Stage CLIP Aggregation	Hierarchical CLIP visual features for robust generated image detection.

📢 Release Status

Item	Status	Notes
🧪 Testing / evaluation code	✅ Released	Use `test.py` for benchmark evaluation.
🏆 Pretrained weights	✅ Released	Use released checkpoints for direct testing.
🏋️ Training code	🚧 Coming next	The next update will provide the training pipeline.
⚙️ Training configs	🚧 Coming next	Full training configuration will be released with training code.
📘 Reproducibility details	🚧 Coming next	More instructions will be added in the next update.

📁 Evaluation Data Format

For the current release, only evaluation is required. Each test dataset folder should follow the common real/fake structure:

test_root/
  progan/
    0_real/
    1_fake/
  stylegan/
    0_real/
    1_fake/

Label meaning:

Folder	Class
`0_real`	Real
`1_fake`	Generated / Fake

🏆 Pretrained Weights

Download the released pretrained weights before running evaluation.

Source	Link	Access Code / Notes
百度网盘	Download	提取码: `8pkh`
Google Drive	Download	\

After downloading, place the checkpoint under runs/DoU/:

runs/
  DoU/
    best.pth

Then use --ckpt runs/DoU/best.pth when running evaluation.

⚡ Usage

✅ Main Command

For the current release, run the evaluation command below with the released checkpoint:

python test.py \
  --data-root  xxx\
  --ckpt runs/DoU/best.pth \
  --output result.xlsx \
  --batch-size 256

🔧 Optional Evaluation Commands

The commands below are alternative or advanced evaluation usages. You do not need to run all of them.

Evaluate selected datasets only

python test.py \
  --data-root xxx \
  --datasets progan stylegan biggan \
  --ckpt runs/DoU/best.pth \
  --output result.xlsx

Run evaluation on CPU

python test.py \
  --data-root xxx \
  --ckpt runs/DoU/best.pth \
  --device cpu

Skip missing dataset folders

python test.py \
  --data-root xxx \
  --ckpt runs/DoU/best.pth \
  --skip-missing

Show all command-line options

python test.py --help

🏋️ Training

The training code is the next part to be released.

Planned next update:

Item	Description
Training pipeline	End-to-end training entry point.
Dataset preparation	CSV generation for training and validation.
Configurations	Hyperparameters and training settings.
Reproducibility guide	Instructions for reproducing reported results.

📌 Notes

Released testing code and pretrained weights can already be used for evaluation.
Training code and training data preparation instructions will be added in the next update.
Please check the command-line options before running on a new environment.