Diffusion watermarks can hand back the original image: LRW hides a compressed prompt in the initial noise, so authentication criteria can be set later
Synopsis
Watermarks on diffusion-generated images usually serve only for tracing; once an image is edited, the mark may still be detectable while content changes go unjudged. LRW losslessly compresses the prompt and embeds it with the random seed, decoding metadata and an HMAC authentication tag into the initial noise, so a verifier checks the tag, rebuilds the same noise and regenerates the original image. On Stable Diffusion 2.1 Base, three latent-space watermark instantiations keep complete-message recovery at or above 93% in most conditions under JPEG compression, Gaussian noise, Gaussian blur and median filtering, giving later authentication a reference image.
Interpretation
The watermark no longer only answers whether an image came from a given model; it returns the original generated image to the verifier as a reference for judging whether content was altered. Earlier semantic watermarks bound the mark to a preset semantic representation of the generated content, yet authentication requirements may only be known after generation, as when a face and a logo in one image carry different integrity demands. On 100 images each from MS-COCO and SDP, LRW with PRC, Gaussian Shading and T2SMark reports complete-message recovery under JPEG quality 60, Gaussian noise, Gaussian blur and median filtering; the T2SMark instantiation reaches 100% on all four distortion types on MS-COCO, and on SDP reports 98% for JPEG, 99% for Gaussian noise, 100% for Gaussian blur and 93% for median filtering.
The original image can be restored exactly because rebuilding it needs far less information than its pixels: a prompt, a random seed and a little decoding metadata. A diffusion model generates an image from seeded Gaussian noise and prompt conditioning, and the same model, sampling settings and numerical execution reproduce the same image, so storing the generation inputs replaces a per-image archive. Prompts are compressed losslessly with arithmetic coding using token probabilities from Dream-v0-Base-7B: 593,968 MS-COCO prompts average 73.09 bits and 67,101 SDP prompts average 379.77 bits, and a 1,500-bit budget covers all MS-COCO prompts and 99.83% of SDP prompts.
Authentication tags and image recovery are split into two steps: message authenticity is checked first, then the application chooses its own comparison. Equating a detectable watermark with authentic content lets forged images that keep a valid mark slip past inspection; protecting the recovery information with HMAC means an attacker without the sender's key cannot forge an acceptable recovery message. HMAC-SHA256 is truncated to 64 bits, and when verification attempts under one key are capped, the probability of accepting a forgery is bounded by the attempt count divided by two to the 64th power; experiments run on Stable Diffusion 2.1 Base with 50 sampling steps and guidance scale 3.
Perspective
The demonstrated setting is Stable Diffusion 2.1 Base with 50 sampling steps and guidance scale 3, watermark instantiations PRC, Gaussian Shading and T2SMark, and distortion and editing tests of 100 images each from MS-COCO and SDP.
Exact recovery depends on the same model weights, sampling settings and numerical execution as the original generation; recovery rates vary with the watermarking method and message length, and the Gaussian Shading instantiation falls to 74% and 58% on SDP under Gaussian blur and median filtering.
