Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

CoSDA: Enhancing the Robustness of Inversion-based Generative Image Watermarking Framework

Authors: Han Fang, Kejiang Chen, Zijin Yang, Bosen Cui, Weiming Zhang, Ee-Chien Chang

AAAI 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that Co SDA effectively enhances watermark robustness while maintaining the visual quality of generated images. [...] To assess the robustness of Co SDA, we adopt the evaluation settings from (Yang et al. 2024), encompassing both the detection and traceability scenarios. [...] All results are obtained from the testing on 50 watermarked images generated with randomly sampled prompts from Stable-Diffusion Prompt. [...] Abaltion Study
Researcher Affiliation Academia 1 National University of Singapore 2 University of Science and Technology of China
Pseudocode No The paper describes its methodology and processes using mathematical equations and textual explanations, but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper discusses the use of existing frameworks and models like Stable Diffusion from Hugging Face, Gaussian Shading, and Tree-Rings. However, it does not provide any explicit statement or link for the open-sourcing of the Co SDA methodology developed in this paper.
Open Datasets Yes During inference, we employ prompts from the Stable-Diffusion Prompt(Gustavosta 2022). https://huggingface.co/datasets/Gustavosta/Stable Diffusion-Prompts. Accessed: 2025-01-10.
Dataset Splits No The paper states that 'All results are obtained from the testing on 50 watermarked images generated with randomly sampled prompts from Stable-Diffusion Prompt.' It also mentions training a drift alignment network ΘDA with 'pairs of distorted-benign latent representations.' However, specific training, validation, or test dataset splits (e.g., percentages, exact counts, or predefined splits with citations) for the data used to train ΘDA are not provided.
Hardware Specification Yes All experiments are performed using Py Torch 1.12.1 and a single NVIDIA-A40 GPU.
Software Dependencies Yes All experiments are performed using Py Torch 1.12.1 and a single NVIDIA-A40 GPU.
Experiment Setup Yes The sampling step and guidance scale are set to 50 and 7.5, respectively. [...] For training and testing with ΘDA, we conduct DDIM inversion with condition and 10 steps. [...] Considering both visual quality and inversion error, we set p = 0.8 for subsequent comparison experiments. [...] The network ΘDA consists of one Single-Conv (the concatenation of Conv-BN-Re LU ), three Res-Block s (He et al. 2016) and one Conv block. To train ΘDA, we minimize the difference between ˆzm V and zm T . The loss function L is set as: L = ˆzm V zm T 2 2