Core Concepts
A training-free plug-and-play watermark framework that can flexibly embed diverse watermarks in the latent space of Stable Diffusion models without retraining the models.
Abstract
The content presents a training-free plug-and-play watermark framework for Stable Diffusion (SD) models. The key highlights are:
The framework embeds diverse watermarks in the latent space of SD models without modifying any components of the models. This allows the watermark to be adapted to different versions of SD without retraining.
The watermark is compressed and embedded in one channel of the latent code, minimizing the impact on the generated image quality. Experiments show the watermarked images have high PSNR, SSIM, and low LPIPS compared to the original images.
The watermark can be effectively extracted using a multi-scale UNet-based decoder, achieving high Normalized Correlation (NC) and Character Accuracy (CA) scores. The framework also demonstrates robustness against common attacks like Gaussian blur, noise, and cropping.
The watermark framework exhibits strong generalization capabilities, allowing the watermark model trained on one version of SD to be directly applied to other versions without retraining.
Extensive experiments validate the effectiveness of the proposed framework in terms of watermark invisibility, watermark extraction quality, and image quality, outperforming existing watermarking methods for SD.
Stats
The content does not provide any specific numerical data or statistics. However, it mentions the following key figures:
The size of the generated images is 512 x 512 pixels.
The size of the latent code is originally 64 x 64 x 4, and the watermark representation is embedded in 64 x 64 x 1.
The binary watermark has a size of 256 x 256 pixels.
Quotes
The content does not contain any direct quotes.