InstantSplamp: Fast and Generalizable Stenography Framework for Generative Gaussian Splatting
Authors: Chenxin Li, Hengyu Liu, Zhiwen Fan, Wuyang Li, Yifan Liu, Panwang Pan, Yixuan Yuan
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments across various potential deployment scenarios demonstrate that Instant Splamp strikes an optimal balance between rendering quality and hiding fidelity, as well as between hiding performance and speed. Extensive experiments demonstrate that Instant Splamp achieves a balance between rendering quality, hiding fidelity, and processing speed, reducing watermarking overhead previously up to 30 times the generation time to nearly zero. We train our model on a filtered subset of the Objaverse dataset Deitke et al. (2023c), excluding low-quality 3D models such as partial scans and those without textures. This filtering process results in a final collection of approximately 80K 3D objects. For training, 100 objects are randomly selected, while a separate test set of 100 unseen objects is reserved for evaluation. We render RGBA images from 40 camera views at a resolution of 256 256 for both training and testing. To evaluate the quality of the recovered hidden information, we assess PSNR, SSIM, and LPIPS. All metrics are calculated on the test set and averaged across all scenarios and embedded images Chen et al. (2022); Li et al. (2022a); Liang et al. (2021). Table 1: Quantitative comparison in rendering and hidden information recovery. Table 2: Ablation study on the proposed key components of Instant Splamp. Robustness Analysis As shown in Fig.6, we can observe that Instant Splamp is robust against common perturbations, such as JPEG compression and Gaussian noise. |
| Researcher Affiliation | Collaboration | Chenxin Li1 , Hengyu Liu1 , Zhiwen Fan2, Wuyang Li1, Yifan Liu1, Panwang Pan3, Yixuan Yuan1 1The Chinese University of Hong Kong, 2University of Texas at Austin, 3Byte Dance |
| Pseudocode | No | The paper describes methods and processes using mathematical formulas and descriptive text, but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Project page: https://gaussian-stego.github.io/. |
| Open Datasets | Yes | We train our model on a filtered subset of the Objaverse dataset Deitke et al. (2023c), excluding low-quality 3D models such as partial scans and those without textures. This filtering process results in a final collection of approximately 80K 3D objects. |
| Dataset Splits | Yes | For training, 100 objects are randomly selected, while a separate test set of 100 unseen objects is reserved for evaluation. |
| Hardware Specification | No | The paper mentions training times and efficiency comparisons but does not specify the exact hardware (e.g., GPU models, CPU types) used for the experiments. |
| Software Dependencies | No | Our approach is deployed over an advanced image-to-3DGS model, LGM Tang et al. (2024), which serves as the Gaussian generator. A simple U-Net is utilized as the decoder for hidden information. Upon the Gaussian generation foundation, we fine-tune a Lo RA Hu et al. (2021) for each watermark image. This training routine typically necessitates training over N=100 objects for approximately 30 epochs, which takes around 20 minutes, and subsequently, it can be generalized to other unseen objects. We employ the Adam W optimizer for optimization, with a learning rate of 1e 4. The paper mentions software components like LGM, U-Net, LoRA, and Adam W optimizer, but does not provide specific version numbers for any of them, nor for underlying frameworks like PyTorch or TensorFlow, or Python versions. |
| Experiment Setup | Yes | Upon the Gaussian generation foundation, we fine-tune a Lo RA Hu et al. (2021) for each watermark image. This training routine typically necessitates training over N=100 objects for approximately 30 epochs, which takes around 20 minutes, and subsequently, it can be generalized to other unseen objects. We employ the Adam W optimizer for optimization, with a learning rate of 1e 4. For hyper-parameters in Eq. (3), we set the weight λ1 = 0.3, λ2 = 1, λ3 = 0.1 for all experiments. |