acceptodds
Withdrawn from ICLR 2027

Flow-Syndrome: Keyed Latent-Carrier Watermarking for Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The growing accessibility of realistic video generation creates an urgent need to preserve provenance through distribution and editing. Existing video watermarking methods encode information in output-space perturbations, generation noise, latent representations, or model parameters, each introducing different dependencies on generator access, inversion, or synchronization under temporal edits. We propose Flow-Syndrome, a generation-time watermarking framework that writes keyed code constraints into the latent representation of a frozen video generator and recovers them with a key-independent detector. A secret key determines a permutation and polarity map over detector carriers together with the keyed payload representation, while the detector itself is trained once per generator family and does not depend on the deployment key. During generation, bounded latent guidance drives detector scores toward the target carrier signs while fidelity and temporal-consistency terms limit perceptual distortion. At detection, the received video is re-encoded with the frozen VAE, carrier scores are aggregated over temporal groups, and the keyed mapping is inverted to recover the payload without access to the original video or inversion of the generative sampling process. Training through decode–transform–encode round trips, together with repeated encoding and temporal aggregation, improves robustness to common spatial, compression, and temporal edits. Experiments on both text-to-video and image-to-video models show strong payload recovery under diverse transformations while preserving video quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Flow-Syndrome Video Watermarking

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_vgA5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Flow-Syndrome, a framework for watermarking generated videos through bounded latent optimization of a frozen video generator. A key-agnostic detector measures latent carriers, while a secret key determines carrier permutation, polarity, and payload encoding. Embedding combines detector guidance with fidelity and temporal-consistency objectives. During extraction, the received video is re-encoded using the corresponding VAE. Detector scores are aggregated across temporal groups, followed by keyed inverse mapping and repetition decoding. The detector is trained through a decode–transform–encode pipeline to improve robustness to video processing. Robustness experiments primarily use Wan2.2-T2V, while quality experiments include text-to-video and image-to-video configurations on Wan2.2 and HunyuanVideo. The results show promising recovery of short payloads under temporal transformations and small changes in reported VBench scores, although recovery deteriorates at larger payload sizes.

Critical strengths

S1. Separating detector learning from deployment-time keying provides practical flexibility. The detector is shared across keys, while keyed carrier assignment and polarity determine the embedded code. This supports key rotation without retraining the detector. Extraction requires neither an unwatermarked reference video nor inversion of the generative sampling trajectory, and the generator parameters remain unchanged. These properties are useful for deployment.

S2. The training procedure addresses the processing and re-encoding channel. Decode–transform–encode training accounts for the discrepancy between generation-time latents and those recovered from processed videos. Combined with temporal replication and aggregation, the method achieves promising short-payload recovery under the tested temporal transformations. The small reported VBench changes provide additional evidence that the method can preserve the measured aspects of video quality.

Critical weaknesses

W1. The incremental benefit of the syndrome construction is not established. The paper does not provide controlled ablations of syndrome encoding, temporal grouping, round-trip training, or guidance-generated positive samples. Consequently, it is difficult to attribute the observed robustness to the proposed components. In particular, the decoder regenerates anchors from the key rather than using observed anchor values for error correction. The benefit of this construction over simpler keyed masking and repetition therefore remains unclear. A matched comparison should establish whether syndrome encoding improves payload recovery, watermark-presence detection, or both. This is an important but potentially addressable gap in the evidence for the method’s contribution.

W2. The evaluation provides limited evidence for complete-message recovery and calibrated presence detection. Robustness is evaluated on 20 videos and summarized primarily by average bit accuracy, without complete-message recovery rates or uncertainty estimates. Several baseline comparisons also differ in payload length, generator, resolution, or frame count, making it difficult to isolate method-specific advantages. The presence test raises a separate calibration question. Its score uses a payload estimated from the same received video, so calibration must account for that estimation step and dependencies among detector outputs. It is unclear whether the reported false-positive controls evaluate this complete unknown-payload procedure or agreement with a prespecified codeword. Zero observed false positives is informative only in conjunction with the number of controls and an uncertainty bound. These limitations do not demonstrate that the method fails, but they prevent a confident assessment of its recovery reliability and operational detection performance.

Questions for authors

Q1. Syndrome versus keyed repetition. At matched payload length, carrier count, temporal redundancy, and perturbation budget, how does the syndrome construction compare with simple keyed repetition? Does it improve complete-message recovery, presence detection, or both?

Q2. Calibration of unknown-payload detection. How is the threshold in Sec. 3.6 selected? Do the App. E controls evaluate the full procedure that first estimates the payload, or agreement with a known target codeword? Please report sample counts, false-positive rates with uncertainty bounds, and detection rates at the selected threshold, separately for unwatermarked and wrong-key controls.

Q3. Recovery and comparison protocols. What fraction of attacked videos yields an entirely correct payload? Please provide uncertainty estimates and identify which comparative conclusions remain supported under matched payload lengths and video generation settings.

Q4. Component contributions. Can ablations isolate round-trip training, temporal grouping, and guidance-generated positive samples? In particular, can the effects of temporal grouping be separated from additional repetition, and those of round-trip training from transformation augmentation alone?

Reviewer_WkwSOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Flow-Syndrome, which embeds keyed watermarks into frozen video generators through latent gradient guidance from a compact detector trained independently of the deployment key. Key-derived permutations and polarities assign repeated payload and anchor bits to detector carriers. Extraction uses VAE re-encoding and temporal aggregation. Evaluation comprises 16 transformations on 20 Wan2.2-T2V clips at several payload sizes and four VBench quality dimensions across Wan2.2 and HunyuanVideo T2V/I2V settings.

Critical strengths

The method keeps generator weights frozen and avoids sampling inversion during extraction. These are useful deployment features, although their practical advantage over existing methods remains unverified.

Critical weaknesses
  1. The independent contribution is not convincingly established. Fixed-detector gradient guidance has already been studied in "Guidance Watermarking for Diffusion Models," but the paper does not adequately explain or evaluate the added value of its video-latent design. The syndrome mask can also be absorbed into the keyed polarity map, leaving repetition and anchor checks as the operative mechanisms. Without comparisons against alternatives or core component ablations, the contribution of the proposed encoding and temporal grouping remains unclear.

  2. The experimental evidence is insufficient to establish reliable detection or a robustness–fidelity advantage. The wrong-key test does not establish the false-positive rate of the blind watermark-presence score. Empirical threshold calibration and detection performance at a fixed false-positive rate are not reported. The robustness evaluation is limited to conventional post-processing transformations and does not assess more advanced watermark attacks. Baseline comparisons are not controlled for payload size and visual quality, and some also differ in generator, resolution, and clip length. The I2V and HunyuanVideo results assess quality without evaluating robustness. These gaps leave both detection reliability and the practical advantage of the proposed method insufficiently supported.

Questions for authors
  1. Contribution: What is the main advance over existing video watermarking methods, and what does the syndrome construction add beyond keyed polarity, repetition, and anchors?

  2. Design evidence: How were the hyperparameters selected? Please provide ablations and sensitivity analyses showing which components account for the reported gains.

  3. Experimental validity: Does the claimed robustness-fidelity advantage persist under matched payloads, generation settings, and quality constraints? Please provide controlled comparisons, attack evaluations for I2V and HunyuanVideo, and detection performance at an empirically calibrated false-positive rate. For broader deployment, how robust is the method against adaptive watermark-removal attacks?

Reviewer_zXygOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Flow-Syndrome, a watermarking framework for video generation. The method embeds watermarks into the latent representations of a frozen video generator using gradient-based guidance. It also introduces a key-independent detector, combined with key-based carrier mapping and temporal redundancy, to enable blind watermark extraction after various video transformations.

The overall framework is clear, and the low embedding overhead is a practical advantage. However, I have major concerns about the technical novelty and experimental evaluation. In its current form, I do not think the contributions are strong enough for acceptance at ICLR.

Critical strengths
  1. Low Computational Overhead: The authors report that the additional generation overhead is less than 2%, which is a practical advantage.
  2. Broad Experimental Coverage: The paper evaluates both T2V and I2V generation using Wan2.2 and HunyuanVideo, and tests various video transformations. The method performs well under some temporal attacks.
Critical weaknesses

1. Limited Technical Novelty

The proposed method mainly combines existing techniques without introducing substantial algorithmic innovations. Key-based permutation and polarity mapping are well-established techniques. Latent watermark embedding and gradient-based guidance have also been explored in previous works, such as Guidance Watermarking (ICLR 2026) and GROW (CVPR 2026). In addition, the loss functions used in this paper are quite standard. The authors should clearly explain what is technically new compared with existing methods.

2. Missing Important Baselines

The paper misses several highly relevant video watermarking methods, including SIGMark (ICLR 2026), VideoMark (2025), SPDMark (CVPR 2026), and DiM (2026). In particular, SIGMark and VideoMark have already proposed methods for generative video watermarking and robustness against temporal attacks. Without direct comparisons with these methods, the experimental results are less convincing.

3. Limited Performance Advantages

The proposed method achieves an average bit accuracy of only 91.8% with a 16-bit watermark, which is lower than VideoSeal (256 bits, 96.9%) and VideoShield (512 bits, 93.1%). Moreover, performance drops significantly as the payload size increases, reaching nearly random guessing at 512 bits. The current results do not show a clear advantage over existing methods.

4. Unclear Contribution of the Syndrome Representation

The paper emphasizes the Syndrome representation in both the title and the method. However, Section 3.4 explicitly states that the method does not perform actual syndrome-based error correction. Instead, robustness mainly comes from repeated encoding and temporal aggregation. What additional benefits does the Syndrome representation provide compared with simple repetition coding? The paper lacks experiments to answer this question.

5. Lack of Key Ablation Studies

The paper lacks systematic ablation studies on its main components, including the Syndrome representation, key-based mapping, temporal grouping, and transformation-aware training. Since the method mainly combines existing techniques, these ablation studies are particularly important to show why each component is necessary.

6. Insufficient Robustness Evaluation

The robustness experiments in Table 1 use only 20 videos, which is a relatively small evaluation set. In addition, the paper does not systematically evaluate adaptive watermark removal or forgery attacks. Therefore, the current experiments are not sufficient to fully support the claimed security and robustness of the method.

7. Insufficient Literature Review

The paper cites relatively few relevant works, and the literature review is not comprehensive. Several closely related studies are missing, making it difficult to understand how this work differs from existing methods. Overall, the literature review does not appear to have been carefully prepared.

8. Writing and Citation Errors

The paper contains several obvious writing and citation errors:

  1. Appendix F contains an unresolved citation: "VidSig ?".
  2. VideoSeal appears twice in the reference list.
  3. VideoShield is incorrectly cited as Hu et al. (2025b) instead of Hu et al. (2025a).

These errors suggest that the authors did not carefully check the manuscript and references before submission.

Questions for authors
  1. Compared with existing works such as Guidance Watermarking and GROW, what are the new technical contributions of the proposed gradient-guided watermark embedding method?
  2. Why are highly relevant methods such as SIGMark and VideoMark missing from the comparisons? Can the authors include these baselines?
  3. What advantages does the Syndrome representation offer over simple repetition coding? Can the authors provide ablation studies to support its effectiveness?
  4. How much do temporal grouping and transformation-aware training contribute to the performance? Can the authors provide corresponding ablation studies?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.