LAID: Lightweight AI-Generated Image Detection in Spatial and Spectral Domains

A framework for benchmarking lightweight models on AIGI detection tasks across spatial, spectral, and fusion domains.

📌 Featured Paper

Featured @ PST2025 - https://arxiv.org/abs/2507.05162

News

🆕 [2025-04-28] Official release of LAID repository

🆕 [2025-06-01] Paper accepted to PST2025

🆕 [2025-06-22] Publicized repository, updated README, and added 4 new transformer models to LAID (EdgeNeXt, MobileViT, MobileViTV2, FastViT)

🆕 [2025-12-01] Added arXiv paper link

Collected Methods

method	paper	test code	train code
ShuffleNet	ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices	✅	✅
EdgeNeXt	EdgeNeXt: Efficiently Amalgamated CNN-Transformer Architecture for Mobile Vision Applications	✅	✅
MobileNetV3	Searching for MobileNetV3	✅	✅
MobileViT	MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer	✅	✅
MobileViTV2	Separable Self-attention for Mobile Vision Transformers	✅	✅
MnasNet	MnasNet: Platform-Aware Neural Architecture Search for Mobile	✅	✅
SqueezeNet	SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size	✅	✅
MobileNetV2	MobileNetV2: Inverted Residuals and Linear Bottlenecks	✅	✅
RegNet	RegNet: Self-Regulated Network for Image Classification	✅	✅
FastViT	FastViT: A Fast Hybrid Vision Transformer using Structural Reparameterization	✅	✅
Lađević et al.	Detection of AI-Generated Synthetic Images with a Lightweight CNN	✅	✅
SpottingDiffusion	SpottingDiffusion: using transfer learning to detect latent diffusion model-synthesized images	✅	✅

Setup

1. Download the GenImage dataset (note: the local disk location of the GenImage dataset is flexible, you can specify its path using the --data_dir command-line argument when running train.py. For detailed instructions, refer to the Training section.
2. (Optional) Download pretrained detection model weights

Training

For training, simply run train.py which will automatically subsample your GenImage download and save the subsampled dataset based on the saved_dataset flag.

Training Parameters

Argument	Type	Default	Description
-e, --epochs	int	100	Number of training epochs.
-b, --batch_size	int	64	Batch size for training.
-lr, --learning_rate	float	1e-4	Learning rate for the optimizer.
-wd, --weight_decay	float	0	Weight decay (L2 regularization) for the optimizer.
-d, --data_dir	str	'GenImage/'	Directory containing the GenImage dataset.
--train_image_count	int	100000	Number of images to subsample for training set.
--val_image_count	int	12500	Number of images to subsample to use for validation and test sets.
--saved_dataset	str	'dataset'	Locatiion of subsampled GenImage dataset.
-m, --model	str	Required	Model to train on AIGI detection. Choices: {"ShuffleNet", "EdgeNeXt", "MobileNetV3", "MobileVitV1", "MobileViTV2, "MNASNet", "SqueezeNet", "MobileNetV2", "RegNet", "FastViT", "Ladevic", "Mulki"}.
-dm, --modality	str	Required	Modality of input data (raw RGB (img) or 2D FFT plot ("freq"). Choices: {"img", "freq"}.
-mc, --model_checkpoint	str	None	Path to a previous model checkpoint for continued training (leave as None if you want to train from scratch).
--output_dir	str	Required	Directory where all output model weights and training loss plots will be saved.
--output_model	str	Required	Name of model weight file to save.
--output_plot	str	Required	Name of training loss plot to save.
-c, --cuda	bool	False	Use CUDA supported GPU for training if available.

Sample usage:

python train.py -b 1024 -m "ShuffleNet" -dm "img" --train_image_count 100000 --val_image_count 16000 --output_dir "outputs" --output_model "chk.pth"  --output_plot "chk_plot.png" -c True

Testing

For test, simply run test.py.

Testing Parameters

Argument	Type	Default	Description
-m, --models_dir	str	Required	Model to train on AIGI detection. Choices: {"ShuffleNet", "EdgeNeXt", "MobileNetV3", "MobileVitV1", "MobileViTV2, "MNASNet", "SqueezeNet", "MobileNetV2", "RegNet", "FastViT", "Ladevic", "Mulki"}.
--test_data_img	str	Required	Location of subsampled spatial test set (default location: dataset/spec/test)
--test_data_spec	str	Required	Location of subsampled spectral test set (default location: dataset/spec/test)
--attack	bool	True	Test models on adversarial attacks.
-c, --cuda	bool	False	Use CUDA supported GPU for training if available.

Sample usage:

python test.py -c True --test_data_img 'dataset/img/test' --test_data_spec 'dataset/spec/test' -m 'outputs/'