Answers are grounded in this paper and cite their sources.
Deep learning-based image watermarking must balance perceptual transparency and reliable watermark recovery, and this balance becomes harder to optimize as the payload grows. This paper presents an adaptive optimization framework for stable high-capacity CNN/GAN image watermarking, in which the composition of the loss function, the weighting of its components and the regularization strength are adapted to the watermark complexity. The framework is evaluated with a fixed architecture for payloads from 16 to 16,384 bits in a distortion-free channel, using baseline, simplified no-mapper and attention-extended configurations, with paired significance tests and Holm–Bonferroni correction on an independent test set. The mapper reconstruction term is shown to provide no gradient, so its removal leaves the gradient unchanged and reduces the per-epoch training time by up to about 40% at high capacity. The run-to-run variability of adversarial training, estimated from gradient-equivalent runs, reaches a median of 0.66 dB in PSNR and is comparable to the effect of the loss weights, while the selected weights and weight decay depend on the watermark complexity. In a pilot study, an attention module improved transparency and decoding reliability mainly at higher capacities. The study provides practical guidance for the design and evaluation of optimization procedures in high-capacity watermarking.
No takes yet. Share an insight, caveat, or question.
Bistroń et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: