RAIN: Region-Aware Inversion Network for Semantic Watermark Extraction

Summary

This paper targets the computational cost of extracting Gaussian Shading (GS) semantic watermarks from diffusion-generated images, which conventionally requires multi-step DDIM inversion. Building on the observation that GS decoding only needs the recovered latent to fall in the correct Gaussian decision regions — not to reconstruct the exact initial noise — the author formulates extraction as one-step conditional regression and argues via extended rectified flow that predicting the conditional mean (or median) of candidate noises near the high-SNR image endpoint suffices. The proposed RAIN extractor uses two lightweight NAFNet branches, one predicting an image-like anchor and the other an intermediate latent whose noise level is regressed by a small head, combined through DDIM-style algebraic reconstruction; training adds worst-case selection over a finite distortion set. On SD v1.5 ablations (Table 1), adversarial training consistently improves robustness; on SD v2.1-base (Table 3), the method surpasses FARI broadly but is mixed against OSI, which wins bit accuracy on 6 of 9 distortions (salt-and-pepper: 99.71% vs 93.11%). Cost drops to 14.394 GFLOPs and 15.468M parameters (~47x fewer than both baselines). Key limitations: the sufficiency assumption remains a stated hypothesis, evaluation covers one GS configuration and two models, and the latency gain depends on two-GPU branch parallelism.