DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection of Diffusion Generated Images
Authors: Baoying Chen, Jishen Zeng, Jianquan Yang, Rui Yang
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results show that detectors enhanced with DRCT achieve over a 10% accuracy improvement in cross-set tests. |
| Researcher Affiliation | Collaboration | 1Alibaba Group 2School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China. |
| Pseudocode | No | No explicit pseudocode block or algorithm was found in the paper. |
| Open Source Code | No | The code, models, and dataset will soon be available at https://github.com/beibuwandeluori/DRCT. |
| Open Datasets | Yes | The compared methods involved in most comparative experiments are trained on the DRCT-2M dataset (utilizing real images from MSCOCO (Lin et al., 2014)) or the Gen Image (Zhu et al., 2023) dataset, as will be specified. |
| Dataset Splits | No | No explicit training/validation/test dataset splits (e.g., exact percentages or sample counts) were provided, nor were specific predefined splits cited for all datasets used. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments were provided. |
| Software Dependencies | No | The paper mentions software like "diffusers" and "CLIP" but does not specify their version numbers or other key software dependencies with specific versioning. |
| Experiment Setup | Yes | During training, all detectors take input images of size 224 224, and during testing, images larger than 224 224 will be center-cropped. To achieve better robustness against post-processing, a range of data augmentations are conducted during training, including horizontal flipping, Gaussian noise disturbance, Gaussian blurring, random rotation, JPEG compression with random quality, brightness and contrast adjustments, and grid dropout. ...the default value for m is 1.0 in our experiments. ...the default value for λ is 0.3 in our experiments. |