RealVDeblur's One-Step Diffusion Clears Blurry Videos in Real-Time
New generative framework restores sharp videos from motion blur in a single step
Video deblurring remains a challenge due to diverse motion patterns and scarce real-world training data. RealVDeblur tackles this with a physically grounded blur synthesis pipeline built from scene-level 3D Gaussian Splatting assets and high-frame-rate videos, generating realistic camera and object motion blur. The framework then leverages a video diffusion prior for restoration, but with a key modification: it disables temporal compression in the VAE and adopts frame-wise encoding to better handle frame-dependent blur variations.
For practical deployment, the authors distill multi-step diffusion into an efficient one-step generator, making inference fast enough for long videos. A training-free Temporal Window Mask stabilizes outputs beyond the training horizon while maintaining constant memory usage. Extensive benchmarks show strong perceptual quality, semantic fidelity, and temporal consistency on unseen videos, as well as improved downstream 3D reconstruction under severe motion blur. Code is available on the project page.
- Physically grounded blur synthesis using 3D Gaussian Splatting for realistic training data covering camera and object motion
- Distills multi-step video diffusion into a one-step generator for efficient, real-time deblurring
- Training-free Temporal Window Mask stabilizes long video inference with constant memory usage
Why It Matters
Enables high-quality video deblurring on mobile devices and for 3D reconstruction under motion blur.