Research map / Video / Video generation
Safe-Sora: Safe Text-to-Video Generation via Graphical Watermarking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Safe-Sora embeds graphical (logo) watermarks into text-to-video generation outputs using coarse-to-fine adaptive patch matching and a 3D wavelet-enhanced Mamba (SFMamba) embedder/extractor.
Problem
Copyright protection for AI-generated videos: existing generative video watermarks embed low-capacity bitstrings, and no method embeds graphical watermarks (logos) into video generation.
Approach
Match each watermark patch to the most visually similar video frame and spatial region in latent space (coarse-to-fine), then embed and recover it with a spatial-frequency Mamba network whose 3D wavelet branch uses a spatiotemporal local scanning order over the eight subbands.
Key results
- Achieves watermark recovery PSNR 37.71 and video PSNR 42.50 versus best baseline PUSNet (28.86 watermark PSNR, 29.98 video PSNR).
- Reports FVD 3.77 against second-best PUSNet at 154.35.
- Robustness evaluated under random erasing, Gaussian blur/noise, rotation, and H.264 compression (CRF 24), with the method best across all attacks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Safe-Sora instead of PUSNet | VideoCrafter2 watermarking · Panda-70M videos | FVD: 154.35 -> 3.77 | Helped |
| Use Safe-Sora instead of Wengnet | VideoCrafter2 watermarking · Panda-70M videos | watermark PSNR: 33.18 -> 37.71 | Helped |
| Remove Coarse-to-Fine Adaptive Patch Matching (ablation) | Safe-Sora · Panda-70M videos | watermark PSNR: 36.71 -> 37.71 | Hurt |
| Replace Routing by Latent with direct pixel-frame similarity (ablation) | Safe-Sora · Panda-70M videos | watermark PSNR: 36.36 -> 37.71 | Hurt |
| Remove Fine Stage spatial localization (ablation) | Safe-Sora · Panda-70M videos | watermark PSNR: 36.88 -> 37.71 | Hurt |
| Replace Spatiotemporal Local Scanning with vanilla 3D scanning (ablation) | Safe-Sora · Panda-70M videos | FVD: 13.16 -> 3.77 | Hurt |
| Use temporal-first instead of spatial-first scanning within each subband (ablation) | Safe-Sora · Panda-70M videos | watermark PSNR: 36.41 -> 37.71 | Hurt |
| Use temporal-first instead of spatial-first scanning within each subband (ablation) | Safe-Sora · Panda-70M videos | tLP: 0.38 -> 0.24 | Helped |
| Remove Multi-Scale Feature Injection (ablation) | Safe-Sora · Panda-70M videos | FVD: 14.11 -> 3.77 | Hurt |
Surprising findings
- The temporal-first scanning variant (w/o SFS) achieves slightly better tLP (0.24 vs 0.38) and LPIPS than the full method.
Limitations
- Limited to static image-based graphical watermarks such as logos or icons; animated or temporally dynamic watermarks are not handled.
- Watermark fidelity slightly lower on Open-Sora backbone (PSNR 35.42) than VideoCrafter2 (37.71).
Experimental setup
Datasets: Panda-70M Logo-2K VidProm ImageNet
Benchmarks: VidProm (100 prompts) Logo-2K watermark test
Metrics: PSNR MAE RMSE SSIM LPIPS tLP FVD
Baselines: Balujanet UDH PUSNet Safe-SD WengNet
Builds on: Mamba DVMark VideoCrafter2 FreqMamba
Closest papers by meaning
- VideoShield: Regulating Diffusion-based Video Generation Models via Watermarking (ICLR 2025 · similarity 0.88)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.87)
- BitMark: Watermarking Bitwise Autoregressive Image Generative Models (NeurIPS 2025 · similarity 0.87)
- Secure Seed-Based Multi-bit Watermarking for Diffusion Models from First Principles (NeurIPS 2026 · similarity 0.86)
- WMarkGPT: Watermarked Image Understanding via Multimodal Large Language Models (ICML 2025 · similarity 0.86)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.