Official code for LaRE2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection (CVPR 2024). In this paper, we firstly propose the reconstruction error in latent space for generated-image detection. Compared with the existing method, we remarkably reduce the cost of feature extraction while preserving the essential information required for the detection of diffusion-generated images.
For LaRE extraction, please refer to DIFT. For model training, please refer to LASTED
We use GenImage as our dataset for training and evaluation. Please refer to this repo for the dataset. After downloading the dataset, we create several annotation files for data loading.
Here is an example of annotation/train_sdv5.txt:
/path/to/dataset/371_sdv5_00145.png 1
/path/to/dataset/n03594945_36929.JPEG 0
...In this context, 1 represents an image generated by AIGC, while 0 signifies a real image.
bash extract_lare.shThe extracted LaRE is stored as a *.pt file, bearing the same name as the input image.
Once LaRE extracted, we need a map_file. It contains all the absolute paths of LaRE. Here is the example:
/path/to/features/3_adm_7.pt 1
/path/to/features/3_adm_34.pt 1
...bash train_classifier_wmap.shbash test.sh- Release code for feature extraction
- Release code for model training