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].
Watermark Anything With Localized Messages
Authors: Tom Sander, Pierre Fernandez, Alain Oliviero Durmus, Teddy Furon, Matthijs Douze
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that WAM is competitive with state-of-the art methods in terms of imperceptibility and robustness, especially against inpainting and splicing, even on high-resolution images. |
| Researcher Affiliation | Collaboration | 1Meta FAIR 2CMAP, Ecole polytechnique 3Centre Inria de L'Universit e de Rennes |
| Pseudocode | No | Section C ALGORITHMS DETAILS describes DBSCAN (Density-Based Spatial Clustering of Applications with Noise) with numbered steps for its functionality but does not present it in a structured pseudocode or algorithm block format, rather as descriptive text. |
| Open Source Code | Yes | Training and inference code and model weights are available at github.com/facebookresearch/watermark-anything. The training and inference code, as well as trained models are available at: https://github. com/facebookresearch/watermark-anything |
| Open Datasets | Yes | We train our model on the MS-COCO training set with blurred faces (Lin et al., 2014)... Table 2b does the same for high-resolution images from the DIV2k (Timofte et al., 2018) validation set. |
| Dataset Splits | Yes | We train our model on the MS-COCO training set... averaged over the first 10k images of the COCO validation set... Table 2b does the same for high-resolution images from the DIV2k (Timofte et al., 2018) validation set. |
| Hardware Specification | Yes | We train with a batch size of 16 per GPU for 300 epochs using 8 V100 GPUs which takes roughly 2 days. |
| Software Dependencies | No | The code is based on Py Torch... The first training phase (Sec. 4.2) is optimized with Adam W (Kingma, 2014; Loshchilov, 2017)... For the extraction of multiple messages, we use the Scikit-learn DBSCAN implementation (Pedregosa et al., 2011). While PyTorch and Scikit-learn are mentioned as key components, specific version numbers for these software components are not explicitly provided in the text. |
| Experiment Setup | Yes | All experiments are run with nbits = 32... We train at resolution h w = 256 256... The first training phase (Sec. 4.2) is optimized with Adam W... with a linear warmup of the learning rate in 5 epochs from 1 10 6 to 1 10 4 and a cosine annealing to 1 10 6. We set λdec = 10, λdet = 1, α = 0.3. We train with a batch size of 16 per GPU for 300 epochs... The second training phase (Sec. 4.3) further trains the model with the JND attenuation for 200 epochs... During this phase, αJND = 2... We use τ = 0.5 to threshold the watermarked pixels, and we choose a rather strict setup with ε = 1 and minsamples = 1000 pixels. |