Image Copy Detection for Diffusion Models

Authors: Wenhao Wang, Yifan Sun, Zhentao Tan, Yi Yang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that PDF-Embedding surpasses protocol-driven methods and non-PDF choices on the D-Rep test set.
Researcher Affiliation Collaboration Wenhao Wang1, Yifan Sun2 , Zhentao Tan2, Yi Yang3 1University of Technology Sydney 2Baidu Inc. 3Zhejiang University
Pseudocode No The paper describes its method with mathematical formulas and textual explanations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The project is publicly available at https://icdiff.github.io/.
Open Datasets Yes A Diffusion Replication (D-Rep) dataset. D-Rep consists of 40, 000 image-replica pairs, in which each replica is generated by a diffusion model. Specifically, the images are from LAION-Aesthetic V2 [1], while their replicas are generated by Stable Diffusion V1.5 [12]. The project is publicly available at https://icdiff.github.io/.
Dataset Splits No The paper states: 'We divide D-Rep into a training set with 90% (36, 000) pairs and a test set with the remaining 10% (4, 000) pairs.' It mentions a training set and a test set but does not explicitly specify a separate validation set split or its size.
Hardware Specification Yes We implement our PDF-Embedding using Py Torch [38] and distribute its training over 8 A100 GPUs.
Software Dependencies No The paper mentions 'Py Torch [38]' and uses 'Vi T-B/16 [37]' as backbone and 'Dei T [40]' for pre-training. However, it does not specify version numbers for any of these software components.
Experiment Setup Yes We resize images to a resolution of 224 224 before training. A batch size of 512 is used, and the total training epochs is 25 with a cosine-decreasing learning rate.