acceptodds
Under review as a conference paper at ICLR 2027

Unifying Watermarking via Dimension-Aware Mapping

Abstract

Deep watermarking methods often share similar encoder–decoder architectures, yet differ substantially in their functional behaviors. We propose **DiM**, a new multi-dimensional watermarking framework that formulates watermarking as **a dimension-aware mapping problem**, thereby unifying existing watermarking methods at the functional level. Under DiM, watermark information is modeled as payloads of different dimensionalities, including one-dimensional binary messages, two-dimensional spatial masks, and three-dimensional spatiotemporal structures. We find that the dimensional configuration of embedding and extraction largely determines the resulting watermarking behavior. Same-dimensional mappings preserve payload structure and support fine-grained control, while cross-dimensional mappings enable spatial or spatiotemporal localization. We instantiate DiM in the video domain, where spatiotemporal representations enable a broader set of dimension mappings. Experiments demonstrate that varying only the embedding and extraction dimensions and their handling modules leads to different watermarking capabilities, including spatiotemporal tamper localization, local embedding control, and recovery of temporal order under frame disruptions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Digital Watermarking, Multi-dimensional Information Embedding, Video Watermarking, Tamper Detection, Copyright Protection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_x8wEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introducrd DiM, a dimension-aware framework that categorizes deep watermarking methods by encoding and extraction dimensionalities (1D, 2D, and 3D). The authors instantiate DiM in video watermarking to support local embedding, tamper localization, and temporal order recovery. Experiments compare different dimensional configurations and baselines across spatial, temporal, and compression distortions.

Critical strengths
  1. Novel perspective: This paper provides an novel perspective for organizing watermarking functionalities through embedding and extraction dimensionalities. The insight that cross-dimensional mappings can support different function is interesting. The framework is intuitive and has potential value as a design guideline.

  2. Comprehensive experiments. The experimental evaluation is comprehensive, covering global and local watermarking, tamper localization, visual quality, and robustness under various distortions, which show the effectiveness of this paper.

Critical weaknesses
  1. Unclear contribution of dimensionality versus richer supervision: Moving from 1D to 2D/3D mappings also introduces additional structural information and stronger localization supervision. Therefore, it is unclear whether the observed improvements stem from the proposed dimension-aware principle or simply from the additional task-specific information. Controlled ablation studies that separate these factors would help clarify the contribution of dimensionality itself.

  2. Incomplete validation of temporal order recovery: As described in Appendix C, long videos are divided into 8-frame clips for processing. However, the paper does not explain how temporal ordering is preserved or recovered across different clips. For example, if a 128-frame video is divided into 16 clips, recovering the correct frame order within each clip does not necessarily recover the original order of the 16 clips. Can the proposed method handle such cross-clip permutations? If so, how is this achieved, and what is the corresponding sequence-level recovery accuracy?

  3. Incomplete computational cost analysis across dimensional mappings: Although Table 6 reports inference throughput (FPS), the computational costs of different dimensional mappings remain unclear. Since these configurations use different network components, their model sizes and computational requirements may differ significantly. Could the authors report the parameter counts, inference time for each mapping? This would help assess whether the comparisons are computationally fair.

Questions for authors

See weaknesses.

3 comments

Dear Reviewer x8wE, thank you very much for your careful review of our paper and constructive comments. We are encouraged by your positive comments on our novel and intuitive perspective, its potential value as a watermarking design guideline, and our comprehensive evaluation of watermarking functionality, visual quality, and robustness. We hope the following responses can clarify potential misunderstandings and alleviate your concerns.


W1. Definition of dimensionality and payload representation

A1: There may be a misunderstanding regarding the meaning of dimensionality in DiM. We clarify it from the following two aspects.

1. Dimensionality includes the corresponding payload representation

As stated in the Abstract (Lines 15–17) and Introduction (Lines 45–47), DiM represents watermark information as payloads of different dimensionalities, including 1D binary messages, 2D spatial masks, and 3D spatiotemporal structures. Therefore, the additional structural information you mention is part of the 3D payload representation itself, rather than an independent addition outside the dimensional mapping.

2. Additional ablation with M1,1\mathcal{M}\\{1,1\\}

We additionally evaluate an M1,1\mathcal{M}\\{1,1\\} baseline using the same MaskWM backbone. Compared with M1,1\mathcal{M}\\{1,1\\}, M1,3\mathcal{M}\\{1,3\\} introduces a 3D payload representation at the extraction stage, thereby realizing a mapping from a low-dimensional space to a high-dimensional space. As shown below, M1,3\mathcal{M}\\{1,3\\} achieves substantially higher extraction accuracy under local embedding, supporting the effectiveness of this mapping.

Mask ratio0–10%11–20%21–30%31–40%41–50%51–60%61–70%71–80%81–90%91–100%
M1,1\mathcal{M}\\{1,1\\} (%)52.6260.0763.3267.4872.8276.5381.4184.6091.1299.67
M1,3\mathcal{M}\\{1,3\\} (%)91.71100.00100.00100.00100.00100.00100.00100.00100.00100.00

W2. Frame-order recovery for long videos

A2:

1. Watermark encoding versus the temporal-protection protocol

Temporal protection has two layers:

  1. Watermark encoding and extraction. This layer embeds and recovers frame-identity information.

  2. Temporal-protection protocol. This layer organizes the codes, authenticates frame identities, and reconstructs temporal order for a particular application.

DiM focuses on the first layer through multi-channel-mask encoding and does not bind the mechanism to a single application protocol. For example, a coarse-grained protocol may assign codes at the 8-frame-clip level, whereas a key-frame protocol may assign distinct codes only to selected frames.

2. Exact sequence recovery on 100-frame videos

To evaluate complete order recovery directly, we use a more challenging frame-wise protection protocol on 100-frame videos.

  • Clip processing. Following Appendix C, each video is processed as a set of 8-frame clips.

  • Within-clip recovery. Multi-channel masks recover frame identities and order within each clip.

  • Across-clip recovery. Because the same within-clip codes may be reused across clips, we use temporal continuity between neighboring frames and clip boundaries, measured by MSE, to infer clip grouping and clip order.

  • Strict success criterion. A video is counted as correct only if all 100 frames are recovered in the exact original order.

Under this strict criterion, the exact sequence-level ordering accuracy is 98.78%.


W3. Parameter counts and inference time

A3:

1. Parameter counts

The parameter counts of each component are summarized below. Across the DiM mappings, the encoder and decoder have the same reported parameter counts, while the mask prediction network differs.

# Params (M)M1,3\mathcal{M}\\{1,3\\} / M2,3\mathcal{M}\\{2,3\\} / M3,3\mathcal{M}\\{3,3\\}M3,2\mathcal{M}\\{3,2\\}
Encoder90.990.9
Decoder91.591.5
Mask Prediction Network3.499.0

2. Inference time

Table 6 already reports inference efficiency in FPS, calculated as the number of processed frames divided by the elapsed time. For clarity, we convert these values into average processing time per frame, using tframe=1000/FPSt_{\mathrm{frame}} = 1000/\mathrm{FPS} in milliseconds.

Time (ms/frame) ↓MaskWMWAMOmniGuardRobust-WideTrustMarkVideoSealRivaGANREVMarkDiM-V
Embedding20.7125.98317.6616.8911.4211.40438.8516.738.72
Extraction52.4435.651615.7754.9728.3010.24407.2711.944.38

Most of my concerns are solved, I have raised my score.

AuthorsOpenReview ↗

Thank You for Your Positive Feedback!

Thank you so much for your positive feedback! It encourages us a lot.

We appreciate your recognition that our rebuttal addressed your main concerns. We fully agree that these clarifications are important, and we will incorporate the additional explanations and results into the revision.

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

This paper proposes DiM, a unified framework that characterizes watermarking through mappings between payloads of different dimensionalities. The authors instantiate the framework in video watermarking through DiM-V, adapting existing encoder–decoder architectures to support global watermark recovery, local embedding, spatiotemporal localization, and temporal-order recovery. Experiments on SA-V compare different mapping configurations and existing watermarking methods under several distortions. Additional evaluations examine multi-channel temporal codes, a second backbone, and high-resolution videos.

Critical strengths

S1: The attempt to organize watermarking functionalities through a common dimensional representation is interesting. It provides a potentially useful perspective for relating global verification, local embedding, and spatiotemporal extraction within a shared design space.

S2: The experimental evaluation is reasonably extensive, covering multiple mapping configurations, nine baseline methods, different distortion categories, and two backbone architectures. The multi-channel frame encoding and evaluations across different mask ratios are also useful aspects of the work.

Critical weaknesses

W1: It is unclear whether the proposed dimensional mapping actually explains the observed watermarking capabilities. The different configurations also change network architectures, mask representations, and prediction procedures. In particular, the high-to-low mapping produces frame-wise 2D masks that collectively still represent a spatiotemporal volume. Without better-controlled experiments, it is difficult to determine whether the reported behaviors arise from the proposed dimensional formulation or from the associated architectural and training choices.

W2: The experimental evidence does not fully support the claimed localization and temporal-order recovery capabilities. The localization protocol relies on predefined watermark-region masks rather than independently manipulated video content, while temporal recovery is primarily evaluated through frame-code accuracy instead of actual sequence reconstruction. The use of the embedding mask during extraction also requires clarification, as it may affect the interpretation of the reported results.

Questions for authors
  1. In Sections 3.2 and 4.2, the distinction between different dimensional mappings is not entirely convincing. For M{3,2}, the decoder predicts frame-wise 2D masks that collectively still form a spatiotemporal representation. Similarly, M{1,3} predicts a spatial watermark-presence map rather than recovering an explicitly embedded 3D payload. How are these mappings fundamentally different from selecting different prediction outputs or network architectures? The authors should clarify what nontrivial behavior the formulation predicts beyond the capabilities explicitly learned during training.
  2. The comparisons do not isolate the effects of dimensional mapping from other design choices. Appendix B.2 uses different 2D and 3D prediction networks, while Appendix A.1 specifies different training resolutions for some configurations. Moreover, there is no M{1,1} model trained with the same backbone, noise layer, JND module, and training schedule. Could the authors provide a matched M{1,1} baseline, including REV{1,1}, to determine whether the reported improvements come from dimensional mapping rather than architectural or training differences?
  3. In Appendix B.1, Eq. (6), the decoder input is constructed by multiplying the received video by the embedding mask M. Is the ground-truth mask used during evaluation, or is it replaced by the predicted mask? If the true mask is required at test time, the reported extraction performance relies on information that may not be available in a blind verification setting. The authors should clarify the complete inference procedure and explain how the mask is obtained during testing.
  4. In Appendix B.1, Eq. (4), partially watermarked videos are generated by combining watermarked and original content using a predefined mask. The localization results therefore appear to measure recovery of watermark-bearing regions rather than detection of independently manipulated content. Furthermore, the qualitative results in Figures 15–25 show grey regions in the tampered examples, whereas Eq. (4) specifies original content outside the watermark mask. Are these actual evaluation inputs or only visualizations? Results on post-embedding video manipulation, such as object removal, inpainting, or cross-video splicing, would be more convincing.
  5. Fig. 6 does not demonstrate complete temporal-order recovery, which reports bit accuracy rather than successful reconstruction of complete frame permutations. The four-channel temporal encoding provides only 15 distinct nonzero identifiers, while Appendix C evaluates videos with more than 100 frames using eight-frame clips. How are frame identities distinguished across different clips? The authors should report exact sequence-recovery accuracy under temporal shuffling and compare against a straightforward baseline that embeds frame indices into individual frames.
  6. I have several concerns about the baseline comparison. Table 2 reports exactly 96.88% frame-level accuracy for several different watermarking methods, while Appendix A.2 uses all-white frames for replacement and insertion attacks. How are these introduced frames scored, and are frame-wise predictions aggregated consistently across the methods? The comparisons also involve different payload lengths and image qualities, making the robustness advantage difficult to interpret without a matched evaluation. Additionally, V2A-Mark should be discussed more explicitly as relevant prior work on video manipulation localization.
  7. Some implementation and distortion details appear inconsistent with the reported results. Appendix B.2 describes a linear message translator whose stated output dimensions would imply an extremely large parameter count if implemented directly. Could the authors clarify the actual layer dimensions and parameter count? Appendix A.2 also specifies Gaussian blur with a kernel size of 1, which would normally leave the image unchanged. Please confirm whether these settings reflect the implementation or are reporting errors.
9 comments
AuthorsOpenReview ↗

Response to Question 1

Dear Reviewer CrwL, thank you very much for your careful review of our paper and detailed technical comments. We are encouraged by your recognition of the potentially useful perspective provided by our dimensional formulation, the breadth of our experimental evaluation, and the usefulness of our multi-channel frame encoding. We hope the following responses can clarify potential misunderstandings and alleviate your concerns

Since the concerns raised in the Weaknesses section are also reflected in the Questions, we address them together in our responses to the Questions below.

Q1. Definition and contribution of dimensional mappings

A1: Since this question involves several aspects, we address them point by point below.

1. Different mappings guide different functional designs

  • M1,3\mathcal{M}\\{1,3\\}: Input: a 1D message. Output: the recovered 1D message plus a 3D spatiotemporal watermark-presence mask. It supports global embedding, local extraction, and tamper localization.

  • M2,3\mathcal{M}\\{2,3\\}: Input: a 1D message plus a 2D spatial mask. Output: the recovered message plus a 3D spatiotemporal mask. It supports spatially local embedding, local extraction, and tamper localization.

  • M3,3\mathcal{M}\\{3,3\\}: Input: a 1D message plus a 3D spatiotemporal mask. Output: the recovered message plus a 3D spatiotemporal mask. It supports spatiotemporally local embedding, local extraction, and tamper localization.

  • M3,2\mathcal{M}\\{3,2\\}: Input: a 1D message plus a 3D multi-channel mask. Output: the recovered message plus frame-wise 2D masks. It supports spatiotemporally local embedding, local extraction, tamper localization, and frame-order recovery.

2. M3,2\mathcal{M}\\{3,2\\} versus M3,3\mathcal{M}\\{3,3\\}

We agree that stacking the frame-wise 2D outputs over time yields a 3D tensor. However, DiM defines the extraction dimension by the elementary prediction operation, not by the dimensionality of the tensor obtained after concatenating all predictions.

  • M3,3\mathcal{M}\\{3,3\\} performs a joint 3D spatiotemporal prediction.

  • M3,2\mathcal{M}\\{3,2\\} decomposes the same spatiotemporal extraction problem into independent frame-wise 2D predictions.

The high-to-low design is introduced because joint 3D prediction becomes less reliable when cross-frame mask correlation is weak. Thus, the distinction lies in the extraction mechanism and its robustness/efficiency trade-off, as discussed in Section 4.2.

3. Clarification of M1,3\mathcal{M}\\{1,3\\}

For M1,3\mathcal{M}\\{1,3\\}, only a 1D global message is embedded. At extraction time, the model predicts a 3D spatiotemporal presence distribution that localizes where the watermark remains detectable. The decoder does not recover a pre-embedded 3D payload. This is precisely the low-to-high setting: low-dimensional watermark information is mapped to a high-dimensional localization output.

4. Relationship to network architectures and training

Implementation versus contribution. Network architectures and prediction procedures are the concrete mechanisms used to realize a mapping; they are not the conceptual contribution of DiM. DiM specifies the input/output representation and functional objective without requiring a particular backbone. Under this guidance, both MaskWM and REVMark exhibit the functions associated with the selected mappings.

DiM is prescriptive, not post-hoc. We do not claim that unseen functions emerge without supervision. DiM guides which representation, prediction target, and supervision should be used; training then realizes that design. For example, original REVMark supports only global embedding and extraction, whereas the DiM-guided variants in Appendix E.1 extend it to local embedding, local extraction, and tamper localization.

AuthorsOpenReview ↗

Response to Question 2-3

Q2. Controlled comparisons across dimensional mappings

A2: Thank you for requesting a more controlled comparison. We agree that the evidence is clearest when cross-method benchmarking, within-backbone controls, and cross-backbone validation are separated. We clarify each below.

1. Network architectures and mask-prediction modules

The architecture changes across mappings are confined primarily to the mask-prediction module, whose form is dictated by the extraction dimension. In particular, when cross-frame correlations in a multi-channel mask are weak, direct joint 3D prediction degrades. M3,2\mathcal{M}\\{3,2\\} therefore decomposes 3D extraction into frame-wise 2D prediction and uses a 2D mask-prediction network. This module implements the defined 2D extraction operation rather than adding an unrelated functional component.

2. Different training resolutions

As stated in Appendix A.1, only the multi-channel-mask experiments reduce the resolution from 256×256 to 128×128 to manage the memory and computation required by multi-channel training. This choice may affect quantitative extraction performance, but it does not change the type of function supported by the mapping.

3. Matched backbone and training settings

As detailed in Appendices B.1–B.2, the mappings share the same base encoder–decoder architecture, noise layer, JND module, and loss-function form. Appendix A.1 also specifies a common optimizer, learning rate, and number of training steps. The mask-prediction module is adjusted only when required by the extraction dimension.

4. REV1,1\mathrm{REV}\\{1,1\\} and cross-backbone validation

The original REVMark is already the REV1,1\mathrm{REV}\\{1,1\\} baseline. Table 1 classifies it as M1,1\mathcal{M}\\{1,1\\}, and Table 2 reports its original results. The purpose of Appendix E.1 is different: it tests whether DiM’s design principle transfers to a second backbone. Applying the relevant mappings to REVMark extends it from global embedding/extraction to local embedding, local extraction, and tamper localization.

5. Additional matched M1,1\mathcal{M}\\{1,1\\} baseline

To address the reviewer’s request directly, we add an M1,1\mathcal{M}\\{1,1\\} baseline built on the same MaskWM backbone and trained under the same general settings. This enables a more controlled comparison of the functions supported by different mappings.

The matched MaskWM results (%) across different mask-ratio ranges are reported below.

Mask ratioM1,1\mathcal{M}\\{1,1\\}M1,3\mathcal{M}\\{1,3\\}M2,3\mathcal{M}\\{2,3\\}M3,3\mathcal{M}\\{3,3\\}
0–10%52.62%91.71%86.87%93.79%
11–20%60.07%100.00%100.00%100.00%
21–30%63.32%100.00%100.00%100.00%
31–40%67.48%100.00%100.00%100.00%
41–50%72.82%100.00%100.00%100.00%
51–60%76.53%100.00%100.00%100.00%
61–70%81.41%100.00%100.00%100.00%
71–80%84.60%100.00%100.00%100.00%
81–90%91.12%100.00%100.00%100.00%
91–100%99.67%100.00%100.00%100.00%

As shown in the table, M1,1\mathcal{M}\\{1,1\\} achieves lower accuracy in local embedding and extraction than the other dimensional mappings, demonstrating their effectiveness in this setting.


Q3. Does Eq. (6) use the ground-truth mask at test time?

1. Training-stage use of the ground-truth mask

A3: No. Eq. (6) describes training. Appendix B is explicitly titled “Training,” and the ground-truth mask is used only to construct the decoder input and supervision during training.

2. Inference uses the predicted mask

At test time, the mask-prediction network first estimates the mask from the received video. The message decoder then performs extraction using this predicted mask, not the ground-truth mask. Therefore, inference does not require knowledge of the embedding location and remains a blind-verification procedure.

3. Role of ground truth in evaluation

Figures 3 and 4 compare the predicted mask with the ground-truth mask only to compute IoU/mIoU. The ground-truth mask is not provided to the model during inference.

4. Revision

We will explicitly separate the training and inference pipelines in the revised appendix and add the full test-time data flow to eliminate this ambiguity.


AuthorsOpenReview ↗

Response to Question 4-5

Q4. Does the localization evaluation use independent post-embedding tampering?

1. Role of Eq. (4)

A4: Yes. The localization results are computed after an independent post-embedding manipulation. Eq. (4) appears in Appendix B (“Training”) and constructs supervision that teaches the model to identify regions in which the watermark remains present; it does not describe the complete evaluation pipeline.

2. Grey regions in Figures 15–25

The grey regions are actual inputs. They are not visualization-only overlays. After watermark embedding, selected regions are replaced with grey content to simulate a local post-embedding manipulation. The model receives this modified video and must infer which regions still contain a valid watermark; missing-watermark regions identify the manipulation.

3. Evaluation procedure

The evaluation pipeline is:

Original video → watermark embedding → post-embedding video manipulation → watermark-region prediction → localization mIoU.

Thus, the reported localization is evaluated after the watermarked video has been modified. We will revise the text and captions to make the distinction between the training construction in Eq. (4) and the test-time tampering pipeline explicit.

Realistic prompt-based manipulation. We have recorded a video demonstrating a realistic prompt-based manipulation scenario, available on anonymous website: https://anonymous.4open.science/r/DIM-FB70/. The white regions indicate the tampered areas. This demonstrates that our model identifies tampered regions without an input mask or prior knowledge of their locations.


Q5. Complete temporal-order recovery

A5: For clarity, we address the questions about Figure 6 and Appendix C separately.

1. Figure 6: Why report mask Bit Accuracy?

1.1 Mask Bit Accuracy

The multi-channel-mask mechanism supports both local embedding and frame-order recovery, and both functions require accurate recovery of each frame’s mask code. Figure 6 therefore reports mask Bit Accuracy as a shared low-level measure.

1.2 Direct evaluation of frame-order recovery

We fully understand your concern. We have therefore additionally computed exact sequence-level ordering accuracy, where a sequence is considered correct only if all frames are restored to their original order. The results are presented below.

Mask ratio0–10%>10%
Exact sequence-level ordering accuracy (%)62.78100

2. Frame-order recovery for long videos

2.1 Watermark encoding versus the temporal-protection protocol

Temporal protection has two layers:

  1. Watermark encoding and extraction. This layer embeds and recovers frame-identity information.

  2. Temporal-protection protocol. This layer organizes the codes, authenticates frame identities, and reconstructs temporal order for a particular application.

DiM focuses on the first layer through multi-channel-mask encoding and does not bind the mechanism to a single application protocol. For example, a coarse-grained protocol may assign codes at the 8-frame-clip level, whereas a key-frame protocol may assign distinct codes only to selected frames.

2.2 Exact sequence recovery on 100-frame videos

To evaluate complete order recovery directly, we use a more challenging frame-wise protection protocol on 100-frame videos.

  • Clip processing. Following Appendix C, each video is processed as a set of 8-frame clips.

  • Within-clip recovery. Multi-channel masks recover frame identities and order within each clip.

  • Across-clip recovery. Because the same within-clip codes may be reused across clips, we use temporal continuity between neighboring frames and clip boundaries, measured by MSE, to infer clip grouping and clip order.

  • Strict success criterion. A video is counted as correct only if all 100 frames are recovered in the exact original order.

Under this strict criterion, the exact sequence-level ordering accuracy is 98.78%.

3. Related methods and the suggested frame-index variant

To our knowledge, no directly comparable post-hoc method jointly supports local embedding and frame-order recovery, and we would be glad to compare against a concrete comparable method if one is suggested.

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

This paper proposes DiM, a framework that represents watermarking as a mapping between payloads of different dimensionalities: one-dimensional binary messages, two-dimensional spatial masks, and three-dimensional spatiotemporal payloads. The authors argue that same-dimensional and cross-dimensional mappings induce different functionalities, including local embedding, tamper localization, and temporal-order recovery. They instantiate the framework as DiM-V for video watermarking using a MaskWM-derived architecture and evaluate several mappings on SA-V videos under visual, geometric, compression, and frame-level distortions. Additional experiments use high-resolution long videos and a REVMark backbone. The reported results show strong robustness and localization performance for DiM-V configurations.

Critical strengths
  1. The dimension-aware formulation provides a useful vocabulary for relating global messages, spatial masks, and spatiotemporal payloads. The mapping table and explicit treatment of low-to-high and high-to-low mappings could help organize future watermarking designs.
  2. The video instantiation is broad and technically interesting. The paper evaluates multiple payload configurations, local extraction, tamper localization, frame-level distortions, high-resolution videos, and an additional REVMark backbone.
Critical weaknesses
  1. The central causal claim—that dimensionality itself determines watermarking functionality—is not sufficiently isolated. Different dimensional settings also change the payload representation, input construction, mask-prediction module, output format, and effective model complexity. The experiments therefore show that complete configurations behave differently, but do not establish that dimensionality alone is responsible.
  2. Several comparisons and capabilities rely on synthetic protocols and potentially non-equivalent baselines. The frame-order recovery experiment assigns explicit binary identity codes to frames, which may be closer to frame identification than general temporal-order recovery. The evaluation also uses short 8-frame, 256×256 clips for the main experiments, while the long-video results rely on clip segmentation and resizing. Reported values near 100% across many distortion categories are not accompanied by variance or statistical uncertainty.
Questions for authors
  1. Can you provide controlled experiments in which architecture, parameter count, training budget, and loss functions are matched across mappings, with only the payload dimensionality changed?
  2. How exactly is the original frame order recovered when frames are duplicated, missing, or visually similar? Please report sequence-level ordering metrics and compare against a simple frame-ID or permutation-classification baseline.
  3. Are the SA-V training and evaluation videos strictly disjoint, and are the baseline methods retrained under the same data, resolution, attack distribution, and compute budget?
  4. Can you report results over multiple training seeds and per-video confidence intervals, especially for the near-perfect robustness and localization numbers in Table 2 and Appendix C?
3 comments

Dear Reviewer rLPQ, thank you very much for your careful review of our paper and thoughtful comments. We are encouraged by your recognition of the value of our dimension-aware formulation, its potential to guide future watermarking designs, and the breadth of our video experiments. We hope the following responses can clarify potential misunderstandings and alleviate your concerns.


W1. Isolating the Role of Dimensionality

R1: Thank you for raising this important concern. We agree that the current experiments do not isolate dimensionality as a single causal variable while holding every architectural detail fixed. However, this is not the claim of DiM. DiM treats the embedding and extraction dimensionalities as defining the payload spaces processed by the model. Realizing a particular mapping therefore necessarily involves the corresponding payload representation, input/output construction, and compatible modules.

1. Payload representation and input/output construction. As stated in the abstract (Lines 15–17) and introduction (Lines 45–47), DiM represents watermark information as 1D messages, 2D spatial masks, or 3D spatiotemporal structures. For example, changing the extraction space from 1D to 3D necessarily changes the prediction target from a global message to a spatiotemporal structure. These changes are intrinsic to the instantiation of the corresponding mapping, rather than independent components introduced outside DiM.

2. Compatible modules and quantitative performance. Different payload spaces require modules capable of processing their corresponding representations. We acknowledge that differences in parameter count and computational complexity may also affect numerical performance, and we therefore do not attribute every quantitative improvement solely to dimensionality. Our claim is instead that the dimensional mapping determines the form of the embedded and extracted information and thereby enables different watermarking functionalities.

Finally, the same mapping–function correspondence is observed on two architecturally distinct backbones, MaskWM and REVMark, indicating that this design principle is not specific to a single architecture.


W2. Protocols, baselines, temporal recovery, and uncertainty

R2: Since this concern involves several distinct aspects of the evaluation, we address the evaluation protocols, baseline fairness, temporal recovery, long-video processing, and statistical uncertainty separately below.

1. Synthetic evaluation protocols

Our robustness evaluation follows protocols commonly used in existing watermarking work, covering four categories of attacks: valuemetric, geometric, frame-level, and compression attacks. These transformations are applied to real videos, and all compared methods are evaluated using the same test videos, attack implementations, and parameters. To improve transparency and reproducibility, we provide the complete evaluation code: https://anonymous.4open.science/r/DIM-FB70/

2. Fairness of baseline comparisons

2.1 Watermark payload length

Why capacities differ. Different baselines have fixed payload capacities determined by their original architectures and released checkpoints. Changing these capacities would require retraining and would no longer evaluate the released methods faithfully. We therefore retain each baseline’s original capacity and evaluate DiM at multiple payload lengths.

  • Table 2. For same-dimensional mappings, the largest baseline payload is 100 bits, while DiM is evaluated at 64, 96, and 128 bits. For cross-dimensional mappings, WAM and MaskWM-D use 32 and 64 bits, respectively, while DiM uses 64 bits.
  • Figure 3. DiM is evaluated at both 64 and 128 bits, covering the capacities of MaskWM-ED (64 bits) and OmniGuard (100 bits).
  • Figure 4. DiM uses 64 bits, matching MaskWM-D (64 bits) and exceeding WAM (32 bits).

Thus, DiM does not obtain an advantage from an artificially smaller payload.

2.2 Network architectures

Faithful comparison of different methods. Following common practice in existing watermarking studies, we evaluate different methods using their original architectures under a unified evaluation protocol. Different watermarking methods naturally use different architectures. Replacing every baseline with a shared network would no longer faithfully evaluate the published methods.

2.3 Evaluation protocols

Matched test conditions. All methods are evaluated using the same test videos, attack categories, and attack parameters. For image watermarking methods, Section 5.1 specifies the frame-wise evaluation and resolution-adaptation procedure used to place them in the same video test environment. These controls make the test conditions as consistent as possible while preserving each method’s original model.

3. Frame-order recovery

3.1 Why Figure 6 reports mask recovery accuracy

The multi-channel mask mechanism supports both local embedding and frame-order recovery, and both functions depend on accurately recovering the masks. We therefore report mask Bit Accuracy in Figure 6 to evaluate this shared requirement, rather than assessing only frame-order recovery.

3.2 Direct evaluation of frame-order recovery

We fully understand your concern. We have therefore additionally computed exact sequence-level ordering accuracy, where a sequence is considered correct only if all frames are restored to their original order. The results are presented below.

Mask ratio0–10%>10%
Exact sequence-level ordering accuracy (%)62.78100

As shown in the table, frame order is recovered with 100% accuracy in the evaluated settings where the mask ratio exceeds 10%. For multi-channel masks with ratios below 10%, complete order recovery remains more challenging. However, mask ratios below 10% are not commonly used in practice.

4. Short clips and long-video evaluation

We would like to clarify that resolution scaling and clip-based processing are standard engineering strategies in watermarking, rather than evaluation choices specific to DiM.

4.1 Resolution scaling

TrustMark explicitly introduces resolution scaling for arbitrary-resolution images, and related adaptation strategies are used by WAM and PixelSeal. The input is processed at the model’s working resolution, after which the predicted residual is restored to the original resolution.

4.2 Long videos

VideoSeal similarly supports chunk-based processing through a configurable chunk_size. Splitting a long video into shorter clips is therefore a common implementation strategy for video watermarking systems.

5. Statistical uncertainty

Following established evaluation protocols in existing watermarking methods, we report average performance across 1,000 test videos for global watermarking and 5,000 for local watermarking, covering diverse attack types and mask ratios. This large-scale evaluation provides a reliable assessment of the methods' overall performance.

AuthorsOpenReview ↗

Response to Questions

Q1. Controlled comparisons across mappings

A1: We clarify the comparison in two parts.

1. Existing mappings already provide controlled comparisons

M2,3\mathcal{M}\\{2,3\\}, and M3,3\mathcal{M}\\{3,3\\} use the same encoder (90.9M parameters), decoder (91.5M), and mask-prediction network (3.4M). They also share the same noise layer, JND module, loss formulation, optimizer, learning rate, batch size of 8, and 200k training steps. Therefore, these mappings compare different embedding payload spaces under the same 3D extraction setting while keeping the architecture, parameter count, and training configuration fixed.

2. M3,2\mathcal{M}\\{3,2\\} is a necessary exception

M3,2\mathcal{M}\\{3,2\\} retains the same encoder, decoder, noise layer, JND module, loss formulation, and training schedule. However, its mask-prediction network differs because its extraction target is different: M3,2\mathcal{M}\\{3,2\\} performs independent frame-wise 2D prediction, whereas the other mappings jointly predict a 3D spatiotemporal structure.

We acknowledge that these implementation differences may affect numerical performance and therefore do not attribute every quantitative improvement solely to dimensionality. Our goal is to study how dimensional mappings guide watermarking functions, rather than claim that dimensionality alone accounts for every numerical improvement.


Q2. Frame duplication, missing frames, and visually similar frames

A2: We clarify how DiM-V handles frame duplication, missing frames, and visually similar frames below.

1. Frame duplication.

Each frame has a unique multi-channel mask tied to its original temporal position. When a frame is duplicated, its mask code is duplicated as well. Therefore, multiple identical codes are extracted from the duplicated frames.

2. Missing frames.

If a frame is removed, its corresponding code is absent. We can therefore infer which frame is missing from the codes extracted from the remaining frames.

3. Visually similar frames.

DiM-V does not infer identity from visual differences between adjacent frames. It assigns an independent mask code to each frame, so high visual similarity does not change the frame-identification mechanism.

4. Direct sequence-level evaluation.

We report the experimental results in W2.3.1.

5. Related methods and the suggested frame-index variant.

To our knowledge, no directly comparable post-hoc method jointly supports local embedding and frame-order recovery, and we would be glad to compare against a concrete comparable method if one is suggested.


Q3. Data split and baseline training conditions

A3: We clarify the data split, baseline training conditions, and matched evaluation protocol below.

1. SA-V split.

The SA-V training and evaluation sets are strictly disjoint; no evaluation video is used during training. We will add the following sentence to Page 6, Line 316:

“Evaluation uses two subsets, and the training and evaluation sets are strictly disjoint, with no overlapping videos.”

2. Baseline training.

We use the officially released checkpoints and original settings of the baseline methods rather than retraining every baseline. These methods have different architectures, training procedures, and compute requirements; retraining all of them under a new shared setup could alter their original performance and would no longer evaluate the released methods faithfully.

3. Matched evaluation.

All baselines use the same test videos, attack types, and attack parameters under one evaluation protocol. Methods restricted to images or fixed resolutions are evaluated using the frame-wise and resolution-adaptation procedures described in Section 5.1.


Q4. Multiple seeds and confidence intervals

A4: We clarify our evaluation setting, sample size, and focus on watermarking functionality below.

1. Watermarking differs from stochastic generation.

Our task embeds and recovers predefined watermark payloads in existing videos, rather than generating new content. Watermarking tasks are less sensitive to random seeds than generative tasks. We follow the convention of reporting average performance over test videos, as discussed in W2.5. This reporting convention should not be interpreted as evidence that training seeds have no effect.

2. Large evaluation sample size.

Our results are averaged over 1,000 test videos for global watermarking and 5,000 for local watermarking, reducing the influence of individual test samples rather than relying on a few selected examples.

3. Focus on watermarking functionality.

More importantly, our central contribution concerns how dimensional mappings guide watermarking functionality: they define the payload spaces for embedding and extraction and guide the corresponding functional design, rather than guarantee identical numerical performance across training seeds.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.