Erase the Mark, Preserve the Motion: Temporally Consistent Removal of Video Watermarks

1 Introduction

• Image-domain signal optimization can cause temporal flicker, while generative resynthesis can produce cross-frame structural motion inconsistency.These issues motivate removing watermarks while preserving motion structure and temporal coherence.Source: §1 · p. 2
• The reconstruction jointly processes the complete sequence with a frozen video DiT, reference-video latent, and pose guidance to encourage consistent identity, texture, and background.The adapter constrains body-part trajectories and relative geometry throughout reconstruction.Source: §1 · p. 2
• Average watermark bit accuracy falls to 0.505, and PoseRegen outperforms the strongest baseline by 9.34% on VBench motion-related quality.Experiments cover eleven representative post-hoc and in-generation watermarking schemes.Source: §1 · p. 2
• The paper frames joint watermark-removal effectiveness and motion consistency as a research direction and, to the authors’ knowledge, the first investigation of that joint objective.The reported results indicate broad watermark removal while preserving smooth and coherent motion.Source: §1 · p. 2
• PoseRegen combines a temporally shared yaw transformation with pose-conditioned video reconstruction and trains a lightweight LoRA adapter using full-sequence Flow Matching.The approach aims to preserve motion structure and cross-frame appearance.Source: §1 · p. 2

2 Related Work

• Video diffusion models provide reconstruction priors, but prior inpainting methods typically assume explicit corrupted regions or user-provided masks, unlike invisible watermark removal.Earlier flow-propagation and temporal-attention approaches can accumulate errors and temporal inconsistency for large missing regions.Source: §2 · pp. 2–3
• Watermarking methods include post-hoc processing of completed videos and in-generation embedding in generator latents or denoising.Removal strategies include generative regeneration, controlled reconstruction, next-frame prediction, novel-view synthesis, and spectral-amplitude optimization.Source: §2 · p. 3
• Existing removal studies primarily optimize watermark evasion and visual fidelity without explicitly preserving moving-part geometry and cross-frame appearance.This work combines a temporally shared viewpoint transformation with pose-conditioned reconstruction adapted to retain motion structure.Source: §2 · p. 3

3 Threat Model

• A video owner or generation platform embeds an invisible watermark before publication, and a downstream verifier uses it to establish provenance and ownership.An adversary may process and redistribute the released video offline without changing its high-level semantic content.Source: §3 · p. 3
• The adversary seeks near-random watermark detection or bit accuracy while avoiding noticeable motion distortion, structural drift, and cross-frame flickering.The stated goal is watermark removal or invalidation while preserving the original video’s visual content and motion.Source: §3 · p. 3
• The attack uses the published video and pretrained generation models without watermark payloads, secret keys, embedding models, detector parameters, or detector queries.Effectiveness may decrease with severe occlusion, ambiguous foreground motion, or inaccurate pose estimation; visible-logo removal and valid-message replacement are out of scope.Source: §3 · p. 3

4 Methodology

• The framework jointly reconstructs the complete sequence using ordered pose guidance and a perturbed source-video reference, rather than processing frames independently.The pose pathway describes foreground motion; the reference pathway provides the transformed source sequence.Source: §4 · pp. 3–4
• An identical 8° horizontal viewpoint transformation is applied across frames to disrupt watermark patterns without adding frame-wise camera jitter.The transformed sequence is encoded as a visual reference for reconstruction.Source: §4 · p. 3
• The complete temporally ordered pose sequence is encoded to guide reconstruction along the subject’s evolving geometry and motion trajectory.The pose latent represents temporal evolution, while the reference latent supplies appearance and scene context.Source: §4 · p. 4
• A lightweight Motion Consistency Adapter modifies a frozen video DiT for pose-conditioned reconstruction, while the backbone, text encoder, and VAE remain frozen.Trainable low-rank parameters are inserted into the DiT blocks.Source: §4 · p. 4
• Full-sequence Flow Matching supervision jointly involves all frames and aligns reconstruction with the specified motion trajectory.The ordered pose condition guides foreground geometry and body-part trajectories; the reference-video latent anchors appearance and background details.Source: §4 · p. 4

5.1 Evaluation Setup

• The evaluation uses HunyuanVideo and Wan2.2, 100 videos per generator, eleven watermark schemes, and bit-accuracy and VBench quality measures.Bit accuracy near 0.5 indicates approximately random recovery; the quality measures include temporal, identity, background, motion, and image-fidelity dimensions.Source: §5.1 · pp. 4–5

5.2 Main Results

• After reconstruction, VideoShield’s extraction accuracy falls to 0.4902 and its localization map labels approximately 98.54% of pixels as edited.Before reconstruction, VideoShield’s localization has F1 0.867, IoU 0.765, and AUC 0.918.Source: §5.2 · pp. 5–6
• 0.505 average bit accuracy, comparable to NFPA, while PoseRegen keeps every evaluated watermark below 0.6 across post-hoc and in-generation schemes.All three video watermark schemes reach 0.498–0.503; the lowest residual bit accuracy on a post-hoc scheme is 0.432.Source: §5.2 · p. 5, Table 1 · p. 6
• PoseRegen leads all evaluated VBench 1.0 and 2.0 metrics; motion smoothness is 0.9913 and motion-order is 0.1884, each beating the strongest baseline.Its CLIP-FID is 0.0823, slightly below WatermarkAttacker’s 0.0825.Source: §5.2 · p. 6, Table 2 · p. 7

5.3 Ablations

• The adapter improves all six evaluated dimensions over the frozen raw video DiT, which already produces more coherent sequences than independently reconstructed frames.The adapter adapts the backbone to temporally ordered pose guidance, improving body-part trajectories, cross-frame appearance, and motion structure.Source: Figure 3 · p. 7, §5.3 · p. 7
• Increasing shared yaw from 0° to 8° progressively lowers VINE bit accuracy, but 10° causes a slight rebound; 8° is the selected default.The paper attributes the rebound to increased reference-pose mismatch and reconstruction effort spent repairing appearance and geometry.Source: Figure 4 · p. 7, §5.3 · p. 7
• Reconstruction strength s = 0.4 balances watermark suppression and video quality; s = 0.1 leaves more watermark information, while s = 1 degrades subject and identity consistency.Full reconstruction at s = 1 provides no clear additional watermark-removal gain.Source: §5.3 · p. 7

• PoseRegen combines shared viewpoint transformation with pose-conditioned full-sequence reconstruction using a frozen video DiT and lightweight Motion Consistency Adapter.Experiments show watermark suppression alongside temporal, identity, motion, and visual-fidelity performance.Source: §6 · p. 7
• The results expose vulnerability in current video watermarking systems.Source: §6 · p. 7
• The authors motivate future evaluations that jointly consider watermark robustness and cross-frame video quality.Source: §6 · p. 8

• VBench and VBench-2.0 metrics assess temporal quality, semantic consistency, and motion rationality in reconstructed videos.Source: A Additional Articulated-Motion… · p. 9
• A source-conditioned evaluation tests whether reconstructed videos preserve the articulated motion trajectory of their watermarked inputs.Source: A Additional Articulated-Motion… · p. 9
• The evaluation focuses on generative reconstruction and prediction attacks, including VINE, Stable Signature, and SIGMark across post-hoc and in-generation mechanisms.Bit-accuracy values are taken from HunyuanVideo results in Table 1.Source: A Additional Articulated-Motion… · p. 9

A.1 Source-Conditioned Motion and Fidelity Metrics

• NJTE compares source and reconstructed 2D joint trajectories, while CLIP-FID compares their frame-level CLIP feature distributions.DWPose extracts the joint sequences; a lower NJTE indicates better motion preservation, and a lower CLIP-FID indicates a closer source feature distribution.Source: §A.1 · pp. 9–10
• In the additional motion-fidelity evaluation across three watermarking schemes, PoseRegen had the lowest average bit accuracy among evaluated generative attacks (0.507), close to random message recovery and slightly below NFPA’s 0.509.Source: §A.1 · p. 10
• NJTE is 0.1424 for PoseRegen, below CtrlRegen (0.1446), WatermarkAttacker (0.1503), and NFPA (0.2661).The evaluation compares source and reconstructed joint trajectories after one video-level alignment.Source: §A.1 · p. 10
• PoseRegen achieves Motion Order 0.1884 and Human Identity 0.8395, outperforming evaluated generative baselines, and CLIP-FID 0.0823, slightly below WatermarkAttacker’s 0.0825.Higher Motion Order and Human Identity and lower CLIP-FID indicate better scores.Source: §A.1 · p. 10

B Additional Component Analysis

• The analysis tests whether temporally ordered pose guidance and the Motion Consistency Adapter contribute to motion and appearance preservation.The shared viewpoint transformation’s watermark-removal contribution is analyzed separately in Section 5.3.Source: B Additional Component Analysis · p. 10
• The shared viewpoint transformation is outside this analysis’s focus; its watermark-removal effect is examined in Section 5.3.Source: B Additional Component Analysis · p. 10
• The component analysis focuses on motion- and appearance-preservation effects during reconstruction.Source: B Additional Component Analysis · p. 10

B.1 Effect of Temporally Ordered Pose Guidance

• The ablation distinguishes using the temporal evolution of poses from treating pose maps as an unordered collection of foreground structures.Source: §B.1 · p. 10
• Shuffling the pose sequence lowers Motion Order from 0.1884 to 0.1014 and raises NJTE from 0.1424 to 0.1573; reversing it raises NJTE to 0.2829.The reversed-sequence trajectory error is approximately twice that of the correctly ordered sequence.Source: §B.1 · p. 10
• Shuffling or reversing poses lowers Temporal Flickering from 0.9878 to 0.9591 or 0.9540 and raises CLIP-FID from 0.0823 to 0.0927 or 0.0904.The largest degradations occur in NJTE and Motion Order.Source: §B.1 · p. 10
• Replacing pose features with zeros lowers Motion Order from 0.1884 to 0.1449 and Human Identity from 0.8395 to 0.7592, while raising CLIP-FID from 0.0823 to 0.0893.Temporal Flickering also falls from 0.9878 to 0.9657.Source: §B.1 · p. 10

B.2 Effect of the Motion Consistency Adapter

• The Motion Consistency Adapter improves joint alignment, motion ordering, temporal consistency, identity preservation, and content fidelity.The frozen video DiT already provides strong temporal coherence by jointly processing the complete sequence.Source: §B.2 · p. 10
• The adapter further improves pose-aligned motion reconstruction and cross-frame appearance consistency rather than acting as a generic frame-smoothing module.Source: §B.2 · pp. 10–11

B.3 Functional Roles of the Components

• The shared viewpoint transformation disrupts watermark-bearing patterns without frame-dependent camera jitter; ordered poses guide foreground geometry, while the reference-video latent supplies source appearance and scene information.Source: §B.3 · p. 11
• The Motion Consistency Adapter adapts the frozen DiT to pose-aligned reconstruction, and the components together remove heterogeneous watermarks while preserving motion and cross-frame appearance.Source: §B.3 · p. 11