Skip to content

KEEP

Stabilise face video super-resolution across frames

Visit official site

What KEEP is

KEEP is described as a video face super-resolution framework built on Kalman filtering that keeps a temporally stable face prior through feature propagation. It fuses information from earlier restored frames to guide the current frame and capture consistent face detail.

What you can do with it

  • Raise face recognition accuracy in surveillance video
  • Improve face clarity in older footage
  • Enhance face clarity in videos uploaded to social media

Model notes

  • A VQGAN generator uses an encoder and a decoder
  • A Kalman filtering network spreads temporal information
  • A Kalman gain network fuses observed and predicted states
  • Cross-frame attention (CFA) layers support local consistency

Getting started

  • Visit the official KEEP page for information and code
  • Read the research paper on how the framework works
  • Download and install the required software environment
  • Prepare a face video dataset for super-resolution

Pros & cons

✓ What we like

  • Feature propagation keeps a stable face prior over time
  • Earlier restored frames guide the current frame
  • Cross-frame attention supports local temporal consistency

! What to watch out for

  • Restoring a frame draws on earlier restored frames
  • The described scope is face video super-resolution

FAQ

What principle is it built on?

The material states it is based on Kalman filtering.

Where was it published?

The material states it appeared at ECCV 2024.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review