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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models
Authors: Haoxin Yang, Bangzhen Liu, Xuemiao Xu, Cheng Xu, Yuyang Yu, Zikai Huang, Yi Wang, Shengfeng He
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Stable Guard consistently outperforms state-of-the-art methods in image fidelity, watermark verification, and tampering localization. (Abstract) 4 Experiments (Section 4 title) |
| Researcher Affiliation | Academia | 1South China University of Technology 2Singapore Management University 3Dongguan University of Technology EMAIL EMAIL EMAIL EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Training Process of Stable Guard. Input: Input image X, iteration steps steps, LDM VAE encoder Eµ, LDM VAE decoder Dν. Output: MPW-VAE decoder Dφ, Mo E-GFN Fψ. |
| Open Source Code | Yes | https://github.com/Harxis/Stable Guard The code is available at https://github.com/Harxis/Stable Guard. |
| Open Datasets | Yes | We train Stable Guard on the COCO training set [56] in a fully self-supervised manner, without requiring any manual tampering annotations. For evaluation, watermark extraction is tested on both the COCO test set and a custom text-to-image (T2I) dataset... Semantic masks are generated with SAM [49] |
| Dataset Splits | Yes | We train Stable Guard on the COCO training set [56]... For evaluation, watermark extraction is tested on both the COCO test set... To assess tampering localization, we focus on robustness against AI-generated manipulations using a large-scale AIGC benchmark of 35,000 images 25,000 from COCO and 10,000 from T2I dataset. |
| Hardware Specification | Yes | All experiments are run for 10 epochs on two NVIDIA RTX 4090D GPUs with a batch size of 16. |
| Software Dependencies | No | Stable Guard is implemented in Py Torch, built upon Stable Diffusion 2.11. (The paper mentions PyTorch and Stable Diffusion 2.11, but does not provide specific version numbers for PyTorch or other libraries.) |
| Experiment Setup | Yes | During training, we freeze the base model and optimize only the Watermark Adapter and Mo E-GFN using the Adam optimizer [59] with a learning rate of 1 10 4. All experiments are run for 10 epochs on two NVIDIA RTX 4090D GPUs with a batch size of 16. The tampering localization expert operates on sub-patches of size n = 8. Loss weights are set as λ0 = 0.2 (Eq. (10)), λ1 = 2, and λ2 = 0.5 (Eq. (11)). |