- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
The paper studies watermarks that hide a multi-bit message in the initial noise of a diffusion model. This noise is a long list of numbers sampled from a standard Gaussian distribution. Each message bit is copied into many of these numbers and decides whether they are positive or negative. A secret key then hides this sign pattern by randomly flipping signs or rotating groups of numbers. The paper shows that, if these flips and rotations are truly random and never reused across noise blocks or images, all such watermarks give the model the same kind of noise. As a result, image edits such as JPEG compression damage the noise equally for all of them. What separates a strong watermark from a weak one is therefore only how many message bits it loses from that same damage. Using this view, the paper introduces ESM, which splits the noise into small blocks, gives all numbers in a block the same size, and rotates each block with its own secret rotation. On two Stable Diffusion models, ESM recovers the full 512-bit message after JPEG compression much more often than two prior watermarks, Gaussian Shading and Spherical Watermark.
Critical strengths
- The theory gives a clean picture of watermarking methods that are usually compared only by running them. It explains why Spherical Watermark's shared rotation makes the noise detectably non-Gaussian (Proposition 1), and its formula reproduces ESM's recovery rate from the measured noise of each image (Section 5.3). The authors also test the theory in a creative way. Running one method's decoder on the noise recovered from another method's images gives nearly the same success rate as running it on its own images (Section 5.1).
- After JPEG compression and resizing, ESM recovers most full messages. Under JPEG50, it recovers 114 and 110 of 128 messages, against at most 46 for the prior methods. The ablation (Table 2, Appendix C.3) also explains the importance of each part, with most of the gain coming from giving all numbers in a block the same size.
Critical weaknesses
- The baselines miss some initial-noise watermarks. PRC [1] carries 512-bit messages, and T2SMark [2] tests up to 1024 bits. PRC also proves its noise cannot be distinguished from Gaussian noise for a fixed key, which this paper leaves to future work. The paper also does not compare with other initial-noise watermarks such as WIND [3], SEAL [4], Tree-Ring [5], and Gaussian Shading++ [6]. They can be compared using a standard metric such as detection at a fixed false-positive rate.
- The evaluation is not sufficient to support practical use. Each setting has 128 images from 16 prompts and 8 keys. The edits are JPEG, resize, and crop. Other standard tests such as noise, blur, brightness, and rotation are missing, and diffusion regeneration uses only 32 images on SD1.5.
- The paper is hard to follow. For example, the abstract uses "Gram invariant" and Figure 1 uses "fiber" and "ideal secret frames" before the reader knows what they mean.
References
[1] Sam Gunn, Xuandong Zhao, and Dawn Song. An Undetectable Watermark for Generative Image Models. ICLR 2025.
[2] Jindong Yang, Han Fang, Weiming Zhang, Nenghai Yu, and Kejiang Chen. T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models. NeurIPS 2025.
[3] Kasra Arabi, Benjamin Feuer, R. Teal Witter, Chinmay Hegde, and Niv Cohen. Hidden in the Noise: Two-Stage Robust Watermarking for Images. ICLR 2025.
[4] Kasra Arabi, R. Teal Witter, Chinmay Hegde, and Niv Cohen. SEAL: Semantic Aware Image Watermarking. ICCV 2025.
[5] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-Rings Watermarks: Invisible Fingerprints for Diffusion Images. NeurIPS 2023.
[6] Zijin Yang, Xin Zhang, Kejiang Chen, Kai Zeng, Qiyi Yao, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian Shading++: Rethinking the Realistic Deployment Challenge of Performance-Lossless Image Watermark for Diffusion Models. arXiv:2504.15026, 2025.
Questions for authors
- When the paper applies an idealized error-correcting code to its measured errors, ESM's advantage shrinks, and in one setting it is no longer significant (Appendix C.2). Does ESM still clearly outperform the other methods when a standard error-correcting code is built into the message?
3 comments
Response to Reviewer z3YD (2/2)
Response to Reviewer z3YD (2/2)
Q1. With an idealized ECC, ESM's advantage shrinks and in one setting is not significant (App. C.2). Does ESM still clearly outperform when a standard ECC is built into the message?
Yes. We built a standard Reed–Solomon code into the message and ran generation and inversion on both backbones (R1): RS(128,64), ×16 repetition, the same 512 net bits, standard hard-decision decoding, n=128.
| backbone, channel | ESM+RS | HN+RS | ESM − HN (points) |
|---|---|---|---|
| SD1.5 JPEG50 | 119 | 96 | +18.0 [+6.2, +32.0] |
| SD2.1 JPEG50 | 124 | 107 | +13.3 [+1.6, +28.9] |
| SD1.5 JPEG30 | 116 | 73 | +33.6 [+16.4, +52.3] |
| SD2.1 JPEG30 | 118 | 70 | +37.5 [+18.0, +57.0] |
All four intervals exclude 0, including SD2.1 JPEG50, the backbone and channel of the non-significant C.2 cell. ESM also beats the rotated-HN control, which shares its frames, in all four (R1). Against GS and SW, our extension of App. C.2's main-run RS analysis (validated image by image on the embedded SD1.5 run) keeps every JPEG interval above 0 up to with their native decoders, e.g. ESM − GS +10.9 [+3.1, +21.1] (SD1.5) and +10.9 [+2.3, +21.9] (SD2.1) at JPEG50, . With idealized soft-assisted decoding (the GMD criterion, which a real errors-and-erasures RS decoder matched in every embedded cell), the largest net payload that keeps ≥ 99% JPEG50 exact recovery is 448 bits for ESM vs 128 for HN and 128 for GS on SD1.5, and 320 vs 16 and 256 on SD2.1 (R1); SW never reaches it.
W2. Evaluation is not sufficient: 128 images (16 prompts × 8 keys); only JPEG, resize and crop; no noise, blur, brightness or rotation; regeneration uses only 32 SD1.5 images.
- Noise, blur, brightness, rotation and more (R5). We added 13 new channels on both backbones (n=32 per cell): JPEG70/10, Gaussian noise σ=0.02/0.05, blur r=1/2, median 3/5, brightness ×1.3/×0.7, contrast ×1.3 and rotation 2°/5°.
- Among ESM, HN, GS and SW, no method exceeds ESM's count in any of the 26 cells. Examples (ESM / HN / GS / SW, SD1.5): JPEG10 14 / 4 / 4 / 0; blur r=2 32 / 24 / 19 / 12; noise σ=0.05 25 / 13 / 13 / 13.
- With n=32 several intervals are wide, so we read the table as consistent support. Intervals excluding 0 include blur r=2 vs HN on both backbones and noise σ=0.05 vs HN on SD1.5.
- PRC (SD2.1) trails or ties ESM in 12 of the 13 and is one image ahead on median 3 (all intervals include 0).
- Rotation and the crop are synchronization failures (R5, R6). Rotation defeats every method tested, PRC included (0/32), though ESM keeps the highest bit accuracy (68–72% at 2°). When the paper's crop is pasted back at its original position without rescaling, ESM recovers 123/128 and 125/128 complete messages, against 0/128 for HN and SW. So for ESM the crop as tested fails through rescaling and misalignment, not through the 36% of the image that is lost. Geometric resynchronization is orthogonal to the modulation, and §6 of the revised PDF will scope it explicitly as a synchronization problem.
- Sample size. The ESM, HN, GS and SW comparisons are paired (same key, prompt, payload, Gaussian draw and nonce), with a crossed key × prompt bootstrap, and both prespecified primary intervals exclude 0. The direction of the effect holds across:
- two backbones;
- an independent 1,024-bit study (47/64 vs 14/64);
- the independent embedded-RS executions;
- a re-execution on identical inputs (SD1.5 JPEG50: ESM 114 vs HN 64);
- 15 of 16 key × backbone groups (one tie).
- Regeneration.
- At strength 0.1, ESM keeps 24/32 vs 16/32 (interval [−12.5, 62.5], not significant at n=32), and its bit-accuracy gain, +0.641 [+0.036, +1.532], excludes 0.
- At 0.3, the bit-accuracy difference is significant: +3.326 [+1.312, +5.359], 98.09% vs 94.76%, i.e. about 2.7× fewer bit errors. The regeneration study keeps the submission's SD1.5, n=32-pair design; at that size the bit-level difference already excludes 0 at both strengths, while the paired PSNR, SSIM and LPIPS intervals all contain 0, so the attack distorts both methods alike (App. C.10).
W3. Hard to follow: the abstract uses "Gram invariant", and Fig. 1 uses "fiber" and "ideal secret frames", before they are defined. The revised PDF will fix each instance:
- The abstract will say "the inner products of blocks that share a secret rotation" and "idealized, independent secret rotations".
- Fig. 1's caption will define the fiber ("the set of rotations that map the codeword to the transmitted block, which the decoder cannot tell apart") and ideal frames ((A1)).
- An "ESM at a glance" paragraph, Algorithm 1 and the pipeline figure will come before the theory.
- Every term will be defined at first use, with a notation table (General Response 3/3).
If a point still seems insufficient, please tell us which one, and we will respond to it before the discussion ends. If these results address the weaknesses you listed, we would appreciate it if you would reconsider your rating.
Thank you for the response. It addressed my error-correction concern. However, the new results show that ESM performs about the same as PRC, and I still find the paper somewhat hard to follow. Therefore, I keep my score.
Response to Reviewer z3YD (1/2)
Response to Reviewer z3YD (1/2)
Thank you for the review, and for describing the theory as "a clean picture" of methods "usually compared only by running them" and the transplant test as creative. We ran new experiments on PRC, a built-in Reed–Solomon code, 13 distortions, detection at a fixed FPR and a crop control; the full tables are in the General Response (R1–R8).
W1. Missing baselines: PRC (512 bits; noise provably indistinguishable for a fixed key), T2SMark (up to 1024 bits), WIND, SEAL, Tree-Ring, Gaussian Shading++. Compare, e.g., by detection at a fixed FPR.
PRC: now evaluated, and cited in the revised PDF (official code, 512-bit payload, our pipeline, both backbones, n=128; R4).
Undetectability under a fixed key is no longer left open. New Prop. J.2 (R3): if HMAC-SHA256 and ChaCha20 are PRFs, the key is secret, the Gaussian draw is fresh per image and nonces are unique, then ESM latents for any number of images are computationally indistinguishable from latents built with truly random frames, which in real arithmetic are exactly N(0,I) for every message (Prop. 1(i)); we do not bound floating-point sampler effects. PRC builds undetectability into its code; ESM obtains it from per-image frames keyed by a unique public nonce (transport and verification in R2) and adds an exact recovery law (Thm. 3).
T2SMark (Yang et al., NeurIPS 2025).
Tree-Ring, WIND and SEAL answer a different question.
They answer "is this ours / which index?", not "which 512-bit message?", so they carry no arbitrary payload to compare on. On your suggested metric, detection at a fixed FPR, ESM already reaches the ceiling on clean, JPEG50, JPEG30 and resize (128/128, below), so on these channels a zero-bit scheme can at most tie it while carrying no payload. The revised PDF will discuss them in §2.
Gaussian Shading++ is a deployment layer around Gaussian Shading. It carries a per-image seed in a separate PRC-coded channel next to GS's payload channel and adds signatures for third-party verification. Its modulation is GS's, which our GS baseline implements with the authors' code (App. A.4). It also shows how ESM's per-image nonce can be carried in band.
Detection at a fixed FPR, now reported (R2). New Prop. J.1: under ideal frames, a wrong nonce or an unwatermarked image gives exactly Bin(512,1/2) agreement with any candidate, so thresholds are analytic. For the implementation, 640,000 wrong-nonce decodes per backbone match this (FPR 0.0092–0.0100 against an exact tail of 0.0095). At aggregate FPR 10−3, with up to 106 candidate nonces, ESM is detected in 128/128 images on clean, JPEG50, JPEG30 and resize on both backbones (95% Clopper–Pearson lower bound 97.2%). HN and GS also reach 128/128, so this metric is saturated for the stronger payload watermarks on these channels. Complete 512-bit recovery is the endpoint that separates them, which is why the paper uses it.