SIGMark: Scalable In-Generation Watermark with Blind Extraction for Video Diffusion

Xinjie Zhu; Zijing Zhao; Hui Jin; Qingxiao Guo; Yilong Ma; Yunhao Wang; Xiaobing Guo; Weifeng Zhang

cs.CVMar 3, 2026

SIGMark: Scalable In-Generation Watermark with Blind Extraction for Video Diffusion

Authors: Xinjie Zhu, Zijing Zhao, Hui Jin, Qingxiao Guo, Yilong Ma, Yunhao Wang, Xiaobing Guo, Weifeng Zhang

Organizations: Lenovo Research, Beijing, China

Abstract

Artificial Intelligence Generated Content (AIGC), particularly video generation with diffusion models, has been advanced rapidly. Invisible watermarking is a key technology for protecting AI-generated videos and tracing harmful content, and thus plays a crucial role in AI safety. Beyond post-processing watermarks which inevitably degrade video quality, recent studies have proposed distortion-free in-generation watermarking for video diffusion models. However, existing in-generation approaches are non-blind: they require maintaining all the message-key pairs and performing template-based matching during extraction, which incurs prohibitive computational costs at scale. Moreover, when applied to modern video diffusion models with causal 3D Variational Autoencoders (VAEs), their robustness against temporal disturbance becomes extremely weak. To overcome these challenges, we propose SIGMark, a Scalable In-Generation watermarking framework with blind extraction for video diffusion. To achieve blind-extraction, we propose to generate watermarked initial noise using a Global set of Frame-wise PseudoRandom Coding keys (GF-PRC), reducing the cost of storing large-scale information while preserving noise distribution and diversity for distortion-free watermarking. To enhance robustness, we further design a Segment Group-Ordering module (SGO) tailored to causal 3D VAEs, ensuring robust watermark inversion during extraction under temporal disturbance. Comprehensive experiments on modern diffusion models show that SIGMark achieves very high bit-accuracy during extraction under both temporal and spatial disturbances with minimal overhead, demonstrating its scalability and robustness. Our project is available at https://jeremyzhao1998.github.io/SIGMark-release/.

Figures & tables

Figure 1: (a) Post-processing watermarks: embedding watermarks in pixel-space which inevitably degrades video quality. (b) Existing in-generation methods: maintaining all the message-key pairs for matching, incurring high extraction costs and poor robustness. (c) Our proposed SIGMark: a blind watermarking framework with global frame-wise PRC keys which is inherently scalable.

Figure 2: Overview of our proposed SIGMark. Embedding: We encode the watermark message into the initial latent noise using a Global set of Frame-wise Pseudo-Random Coding (GF-PRC) keys. The diffusion model then denoises this noise into video frames that carry the embedded messages. Extraction: A (possibly disturbed) video is first processed by our proposed Segment Group-Ordering (SGO) module to recover the correct causal frame grouping, then inverted to obtain the latent noise, from which the message is decoded using the GF-PRC keys. The system stores only the GF-PRC keys for both embedding and extraction, enabling blind watermarking.

Figure 3: Watermark embedding with GF-PRC scheme. We set $f_{l}=2,f=8$ with compression ratio $d_{t}=f/f_{l}=4$ as an example. Orange and purple denote message $m[0]$ and $m[1]$ respectively.

Figure 4: Segment Group-Ordering (SGO) module. We set compression ratio $d_{t}=f/f_{l}=4$ as an example. When temporal disturbances (e.g., clipping or frame drops) occur, the causal grouping is disrupted; without re-ordering, this leads to incorrectly encoded latent features. Our SGO module restores the correct grouping and ordering, yielding robust latent features for video inversion.

Diffusion model		HunyuanVideo T2V				HunyuanVideo I2V
Watermarking		512 bits		512 $\times$ 16 bits		512 bits		512 $\times$ 16 bits
Method	Category	Bit acc $\uparrow$	V-score $\uparrow$	Bit acc $\uparrow$	V-score $\uparrow$	Bit acc $\uparrow$	V-score $\uparrow$	Bit acc $\uparrow$	V-score $\uparrow$
No-mark	–	–	0.490	–	0.490	–	0.463	–	0.463
DCT	Post	0.889	0.424	0.862	0.423	0.890	0.452	0.858	0.456
DT-CWT	Post	0.619	0.416	0.650	0.436	0.627	0.458	0.611	0.463
VideoMark	None-blind	0.873	0.507	0.758	0.502	0.846	0.483	0.707	0.482

Table 1: Video watermarking results: message recovery bit accuracy and video quality score. 2 2 2 We implement DCT and DT-CWT through the open-source code of video-invisible-watermark and blind-video-watermark respectively, and we implement VideoMark and VideoShield by adapting their officially released code and hyper-parameter to new diffusion models and prompts.

Method	Spatial disturbance				Temporal disturbance
Method	w/o	G.noise	cmprs	blur	w/o	drop	insert	clip
VideoMark	0.85	0.64 ↓0.21	0.63 ↓0.22	0.64 ↓0.21	0.71	0.52 ↓0.19	0.51 ↓0.20	0.51 ↓0.20
VideoShield	1.00	1.00 ↓0.00	0.99 ↓0.01	1.00 ↓0.00	0.99	0.89 ↓0.10	0.84 ↓0.15	0.83 ↓0.16
SIGMark(Ours)	0.98	0.89 ↓0.09	0.84 ↓0.14	0.95 ↓0.03	0.91	0.81 ↓0.10	0.87 ↓0.04	0.85 ↓0.06

Table 2: Watermark extraction bit accuracy under disturbances on HunyuanVideo I2V.

Method	Watermark embedding		Watermark extraction
Method	Single PRC	GF-PRC(Ours)	w/o SGO	w/o OF-seg	w/o SW-det	SGO(Ours)
Bit acc	0.707	0.905	0.534	0.762	0.823	0.869

Table 3: Ablation study on our proposed modules on HunyuanVideo I2V

Figure 5: The decoding time cost during watermark extraction.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Dimension	Prompt Text	Image
Complex Plot	The race commenced with explosive energy as the first runner from Team A surged ahead, capturing the attention of all spectators. His swift pace set the tone, but during the baton exchange, a slight fumble nearly cost them. Despite the near mishap, he managed to hand off the baton to the second runner, who, with fierce determination, pursued Team B relentlessly, overtaking them on the curve. However, the earlier error meant their lead was slim. As the third runner took over, nerves were palpable, yet he maintained their position, though Team C’s strategic pacing brought them dangerously close. In the final handoff, the pressure was immense. The last runner from Team A accelerated on the penultimate turn, unleashing a powerful sprint that widened the gap. With unwavering focus, he crossed the finish line triumphantly, securing victory for Team A amidst roaring cheers.
Composition	A magnificent lion, its golden mane flowing like a regal crown, possesses the expansive, powerful wings of an eagle, each feather shimmering under the sun’s radiant glow. It soars effortlessly through a vast, azure sky, the clouds parting gracefully in its wake. The lion’s eyes, sharp and focused, scan the horizon with a kingly gaze, while its muscular body glides with the grace of a seasoned aviator. Below, the earth stretches out in a patchwork of greens and browns, rivers glistening like silver ribbons. The scene captures a breathtaking blend of strength and elegance, as the majestic creature commands the heavens with unparalleled ease.
Dynamic Attribute	In a serene forest, the camera captures a mesmerizing transformation as vibrant red leaves slowly transition to lush green. The scene begins with a close-up of a single leaf, its deep crimson hue glowing under the gentle sunlight. As the camera pans out, the surrounding foliage reveals a tapestry of red, orange, and yellow, creating a warm, autumnal atmosphere. Gradually, the colors shift, with hints of green emerging, symbolizing the renewal of life. The sunlight filters through the canopy, casting dappled shadows on the forest floor, while a gentle breeze rustles the leaves, enhancing the sense of change and rebirth.
Dynamic Spatial Relationship	A playful dog, with its fur slightly ruffled by the breeze, stands on the left side of a wooden dining table, its ears perked up and eyes full of curiosity. The table is surrounded by cozy chairs and a vase with fresh flowers on top. As the dog gets excited, it begins to sprint energetically towards the front of the table, its paws making soft thudding sounds against the wooden floor, eager to reach the other side where a chew toy lies waiting.
Instance Preservation	A vibrant orange dog, with a sleek coat and playful demeanor, sprints joyfully across a sunlit meadow, its ears flapping in the breeze. The dog’s eyes sparkle with excitement as it bounds over the lush green grass, leaving a trail of paw prints behind. In the background, wildflowers sway gently, adding bursts of color to the scene. The sun casts a warm glow, highlighting the dog’s energetic movements and the natural beauty surrounding it. As the dog runs, its tail wags enthusiastically, embodying pure joy and freedom in the open, expansive landscape.

Appendix

Table 1: Our constructed prompt set – Sample of 5 dimensions with representative prompts

Dimension	Hunyuan-I2V	Hunyuan-T2V	Wan-I2V	Wan-T2V
complex plot	1.0000	0.9292	1.0000	1.0000
composition	1.0000	0.9313	1.0000	1.0000
dynamic attribute	1.0000	0.8878	1.0000	0.8974
dynamic spatial relationship	0.9773	0.9991	1.0000	0.9554
instance preservation	0.9117	0.9584	1.0000	0.9150
motion rationality	0.9684	1.0000	1.0000	0.9379

Appendix

Table 2: Results of different models across different dimensions

Figure 1: PRC error-tolerance curve and the inversion accuracy of different video diffusion models.

Diffusion Model Watermarking

Diffusion model watermarking embeds detectable signals in generated content or models to support provenance, ownership verification, and attribution.

See all 46 papers

Dec 12, 2024

LVMark: Robust Watermark for Latent Video Diffusion Models

Rapid advancements in video diffusion models have enabled the creation of realistic videos, raising concerns about unauthorized use and driving the demand for techniques to protect model ownership. Existing watermarking methods suffer from two key limitations: they overlook temporal consistency due to conventional watermark decoders and degrade the visual quality of the generated videos. To address these issues, we introduce a robust watermarking method for latent video diffusion models named Latent Video Diffusion Watermarking (LVMark). We propose a novel watermark decoder tailored for generated videos by learning the consistency between adjacent frames. It ensures accurate message decoding, even under malicious attacks, by combining the low-frequency components of the three-dimensional wavelet domain with the color features of the video. Additionally, we train a latent decoder to maintain the visual fidelity of the generated video. Watermarks are embedded into layers with minimal impact on visual appearance using an importance-based weight modulation strategy. We optimize both the watermark decoder and the latent decoder of diffusion model, effectively balancing the trade-off between visual quality and bit accuracy. Our experiments show that our method embeds invisible watermarks into video diffusion models, ensuring robust decoding accuracy with 512-bit capacity, even under distortions.

Youngdong Jang, MinHyuk Jang, JaeHyeok Lee +4

Department of Artificial Intelligence, Korea University, Seoul 02841, Republic of Korea · Google DeepMind, Mountain View, CA 94043, USA

Diffusion Model Watermarking Video Diffusion Models Latent Diffusion Models Robust Watermarking

Ask PDF

Mar 19, 2026

Transferable Multi-Bit Watermarking Across Frozen Diffusion Models via Latent Consistency Bridges

As generative AI advances, global governance frameworks increasingly mandate verifiable content provenance. However, existing watermarking techniques face a critical policy-to-technology disconnect: sampling-based methods require computationally prohibitive inversion, while fine-tuning approaches are tethered to specific model checkpoints, hindering standardized, cross-model oversight. To bridge this gap, we introduce DiffMark, a plug-and-play multi-bit watermarking framework. DiffMark embeds a persistent, learned perturbation into every denoising step of a frozen diffusion model, accumulating a recoverable signal in the final latent space. To enable efficient training through the frozen network, we utilize Latent Consistency Models (LCMs) as a differentiable training bridge. DiffMark achieves 64-bit extraction in a single 16.4 ms forward pass, which is a

45\times

speed-up over inversion baselines. By enabling per-image key flexibility and cross-architecture transferability without retraining, DiffMark provides the practical, scalable technical tooling necessary to operationalize user accountability and enforce emerging AI governance mandates.

Hong-Hanh Nguyen-Le, Van-Tuan Tran, Thuc D. Nguyen +1

School of Computer Science, University College Dublin, Ireland · School of Computer Science and Statistics, Trinity College Dublin, Ireland · Department of Knowledge Engineering, University of Science, VNU-HCMC, Vietnam

Multi-Bit Watermarking Diffusion Model Watermarking

Ask PDF

Feb 27, 2026

SKeDA: A Generative Watermarking Framework for Text-to-video Diffusion Models

The rise of text-to-video generation models has raised growing concerns over content authenticity, copyright protection, and malicious misuse. Watermarking serves as an effective mechanism for regulating such AI-generated content, where high fidelity and strong robustness are particularly critical. Recent generative image watermarking methods provide a promising foundation by leveraging watermark information and pseudo-random keys to control the initial sampling noise, enabling lossless embedding. However, directly extending these techniques to videos introduces two key limitations: Existing designs implicitly rely on strict alignment between video frames and frame-dependent pseudo-random binary sequences used for watermark encryption. Once this alignment is disrupted, subsequent watermark extraction becomes unreliable; and Video-specific distortions, such as inter-frame compression, significantly degrade watermark reliability. To address these issues, we propose SKeDA, a generative watermarking framework tailored for text-to-video diffusion models. SKeDA consists of two components: (1) Shuffle-Key-based Distribution-preserving Sampling (SKe) employs a single base pseudo-random binary sequence for watermark encryption and derives frame-level encryption sequences through permutation. This design transforms watermark extraction from synchronization-sensitive sequence decoding into permutation-tolerant set-level aggregation, substantially improving robustness against frame reordering and loss; and (2) Differential Attention (DA), which computes inter-frame differences and dynamically adjusts attention weights during extraction, enhancing robustness against temporal distortions. Extensive experiments demonstrate that SKeDA preserves high video generation quality and watermark robustness.

Yang Yang, Xinze Zou, Zehua Ma +2

the School of Electronic and Information Engineering, Anhui University, Hefei, Anhui 230601, China · Anhui Province Key Laboratory of Digital Security and the CAS Key Laboratory of Electromagnetic Space Information, University of Science and Technology of China, Hefei 230026, China · the School of Computing, National University of Singapore, Singapore 117417

Diffusion Model Watermarking Robust Watermarking Video Diffusion Models T2V Generation

Ask PDF

Abstract

Diffusion Model Watermarking

Explore similar work

LVMark: Robust Watermark for Latent Video Diffusion Models

Transferable Multi-Bit Watermarking Across Frozen Diffusion Models via Latent Consistency Bridges

SKeDA: A Generative Watermarking Framework for Text-to-video Diffusion Models