FARI: Robust One-Step Inversion for Watermarking in Diffusion Models
Researchers propose FARI, a robust one-step inversion method for watermarking in diffusion models. The approach addresses the bottleneck of slow and error-prone inversion by leveraging two key observations: inversion trajectories have lower curvature than forward generation, making them compressible for low-NFE approximation, and external distortions dominate error in watermark verification, reducing the trade-off between speed and truncation error. This enables faster inversion without sacrificing robustness.
Key facts
- FARI is a one-step inversion method for watermarking in diffusion models.
- Inversion trajectories have lower curvature than forward generation paths.
- External distortions dominate error in watermark verification.
- Existing methods over-optimize internal truncation error.
- The approach aims to meet dual demands of speed and robustness.
- Paper published on arXiv with ID 2607.26723.
Entities
Institutions
- arXiv