Researchers have unveiled a watermarking system that can embed multi‑bit messages into any existing photo without any training or fine‑tuning. The method works by first inverting a diffusion model’s latent vector, then selectively flipping the signs of chosen latent coordinates under a secret key. The changed coordinates are further refined through iterative latent optimization, and finally blended back into the image using a gamma‑mixing step that preserves visual quality.

Because the watermark is stored in the latent space rather than in raw pixels, it remains invisible to the human eye while remaining highly resistant to attacks that try to regenerate the image via diffusion models. On three public datasets—MS‑COCO, DiffusionDB, and WikiArt—the system achieved structural similarity indices above 0.92, peak signal‑to‑noise ratios near 29 dB, and a flexible decoding success rate of 99.9 % for clean images.

When challenged with the latest Stable Diffusion‑based regeneration attack at a quality‑preserving step of t = 0.3, the new framework decoded a full 32‑bit payload on 94.6 % of the 500 MS‑COCO images tested. The three‑layer decoder—combining majority voting, BCH error correction, and Chase decoding—provided the most robust performance, outperforming competing systems by over 10 %.