Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Zero-Shot Video Restoration and Enhancement with Assistance of Video Diffusion Models

Cong Cao, Huanjing Yue, Shangbin Xie, Xin Liu, Jingyu Yang

TL;DR

The paper tackles the problem of temporal instability when applying zero-shot image restoration diffusion methods to videos. It introduces ZVRV, a training-free framework that fuses image-restoration latents with both homologous and heterogeneous video diffusion priors, guided by a chain-of-thought–based fusion ratio strategy and augmented by temporal-strengthening post-processing. Key contributions include the homologous latent fusion, heterogeneous latent fusion, and a CoT-based fusion ratio strategy, along with a temporal refinement step using an I2V diffusion model; together these yield improved temporal consistency and visual quality across zero-shot video super-resolution and low-light enhancement tasks. Experiments on multiple datasets show that ZVRV outperforms state-of-the-art zero-shot and some supervised methods across a suite of metrics, demonstrating the practical impact of leveraging video diffusion priors to stabilize zero-shot video restoration and enhancement without additional training.

Abstract

Although diffusion-based zero-shot image restoration and enhancement methods have achieved great success, applying them to video restoration or enhancement will lead to severe temporal flickering. In this paper, we propose the first framework that utilizes the rapidly-developed video diffusion model to assist the image-based method in maintaining more temporal consistency for zero-shot video restoration and enhancement. We propose homologous latents fusion, heterogenous latents fusion, and a COT-based fusion ratio strategy to utilize both homologous and heterogenous text-to-video diffusion models to complement the image method. Moreover, we propose temporal-strengthening post-processing to utilize the image-to-video diffusion model to further improve temporal consistency. Our method is training-free and can be applied to any diffusion-based image restoration and enhancement methods. Experimental results demonstrate the superiority of the proposed method.

Zero-Shot Video Restoration and Enhancement with Assistance of Video Diffusion Models

TL;DR

The paper tackles the problem of temporal instability when applying zero-shot image restoration diffusion methods to videos. It introduces ZVRV, a training-free framework that fuses image-restoration latents with both homologous and heterogeneous video diffusion priors, guided by a chain-of-thought–based fusion ratio strategy and augmented by temporal-strengthening post-processing. Key contributions include the homologous latent fusion, heterogeneous latent fusion, and a CoT-based fusion ratio strategy, along with a temporal refinement step using an I2V diffusion model; together these yield improved temporal consistency and visual quality across zero-shot video super-resolution and low-light enhancement tasks. Experiments on multiple datasets show that ZVRV outperforms state-of-the-art zero-shot and some supervised methods across a suite of metrics, demonstrating the practical impact of leveraging video diffusion priors to stabilize zero-shot video restoration and enhancement without additional training.

Abstract

Although diffusion-based zero-shot image restoration and enhancement methods have achieved great success, applying them to video restoration or enhancement will lead to severe temporal flickering. In this paper, we propose the first framework that utilizes the rapidly-developed video diffusion model to assist the image-based method in maintaining more temporal consistency for zero-shot video restoration and enhancement. We propose homologous latents fusion, heterogenous latents fusion, and a COT-based fusion ratio strategy to utilize both homologous and heterogenous text-to-video diffusion models to complement the image method. Moreover, we propose temporal-strengthening post-processing to utilize the image-to-video diffusion model to further improve temporal consistency. Our method is training-free and can be applied to any diffusion-based image restoration and enhancement methods. Experimental results demonstrate the superiority of the proposed method.
Paper Structure (19 sections, 17 equations, 5 figures, 5 tables)

This paper contains 19 sections, 17 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Zero-Shot IR with Latent Diffusion Models
  4. Zero-Shot Video Editing
  5. Zero-Shot Video Restoration and Enhancement
  6. Background
  7. Latent Diffusion Models
  8. DDIM Sampling
  9. Latent Diffusion Models for Zero-Shot IR
  10. Method
  11. Homologous Latents Fusion
  12. Heterogenous Latents Fusion
  13. COT-Based Fusion Ratio Strategy
  14. temporal-strengthening post-processing
  15. Experiments
  16. ...and 4 more sections

Figures (5)

  • Figure 1: Overview of our framework.
  • Figure 2: Our COT-Based Fusion Ratio Strategy. We design a process reward model based on video quality as test-time verifiers within CoT reasoning paths.
  • Figure 3: Visual quality comparison for zero-shot 4$\times$ video super-resolution. Zoom in for better observation.
  • Figure 4: Visual quality comparison for zero-shot low-light video enhancement. Zoom in for better observation.
  • Figure 5: Visual quality comparison for 4$\times$ blind video super-resolution. Zoom in for better observation.