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..
Ultra-high Resolution Watermarking Framework Resistant to Extreme Cropping and Scaling
Authors: Nan Sun, LuYu Yuan, Han Fang, Yuxing Lu, Hefei Ling, Sijing Xie, Chengxin Zhao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conduct extensive experiments to evaluate the effectiveness of our proposed INR-based watermarking method. First, we describe the experimental setup. Then, we compare our method with previous SOTA models under low resolution. Subsequently, we evaluate our method across various resolutions against other large-image watermarking approaches, demonstrating its superior performance. Finally, ablation studies assess the contribution of proposed components. |
| Researcher Affiliation | Academia | Nan Sun1, Lu Yu Yuan1, Han Fang2, Yuxing Lu3, Hefei Ling1, Sijing Xie1, Chengxin Zhao1 1School of Computer Science and Technology, Huazhong University of Science and Technology 2National University of Singapore 3Peking University |
| Pseudocode | No | The paper describes the methodology in prose and architectural diagrams (Figure 2), but does not present any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper's main body and supplementary materials (A.3) do not contain an explicit statement or link to the authors' open-source code for the methodology described. |
| Open Datasets | Yes | Our model is trained on the high-resolution DIV2K (Agustsson and Timofte, 2017) image dataset. ... we test it not only on DIV2K but also on 200 separately sampled images from each of the COCO (Lin et al., 2014) and FFHQ (Karras et al., 2019) datasets. |
| Dataset Splits | No | For each training iteration, we randomly select an image from the dataset and apply a random scaling operation, where the scaling factor is chosen from the range [0.06, 1]. Following the scaling, we randomly crop a 128 128 image patch from the scaled image, which serves as the input I to the model. ... We randomly generate training samples, with a training set size of 50,000 samples. ... we test it not only on DIV2K but also on 200 separately sampled images from each of the COCO (Lin et al., 2014) and FFHQ (Karras et al., 2019) datasets. |
| Hardware Specification | Yes | The model is trained using the Adam W (Loshchilov and Hutter, 2017) optimizer with a learning rate of 4 10 4. The batch size is set to 32, and the training is conducted for 2000 epochs across two NVIDIA RTX 3090 24G GPUs. ... We use AMD Ryzen 7 7840HS for our CPU and NVIDIA RTX 3090 24G for our GPU for testing. |
| Software Dependencies | No | The combined noise layer is implemented using Kornia (Riba et al., 2020)... |
| Experiment Setup | Yes | The model is trained using the Adam W (Loshchilov and Hutter, 2017) optimizer with a learning rate of 4 10 4. The batch size is set to 32, and the training is conducted for 2000 epochs across two NVIDIA RTX 3090 24G GPUs. ... To sample submatrices of arbitrary size from ψ, we use a fixed r r coordinate grid C, where r is typically set to 128. ... a set of feature grids with varying resolutions L = {Li}n to represent the embedded features of the watermark template, where n (default 4) denotes the number of feature grids. Each grid Li Rd 2i+4 2i+4 is a learnable parameterized matrix, with d (default 32) representing the dimension of the features. ... Γ( ) is a linear layer used to project the features into the de-dimensional space (default 256). ... Then we randomly sample an r r image block I and obtain the watermarked image as I = I + γW, where γ (default 0.02) controls embedding strength. |