ARTFEED — Contemporary Art Intelligence

AVIS: Autoregressive Diffusion Model for Real-Time Video Restoration

ai-technology · 2026-05-22

A team of researchers has introduced the Autoregressive Video Inverse problem Solver (AVIS), a system that leverages autoregressive video diffusion models to enhance video restoration in real-time. This innovation decreases initial latency from 114 seconds to just 4 seconds and boosts throughput from 0.71 to 1.18 frames per second (FPS). AVIS begins reverse diffusion with a measurement-consistent estimate to minimize the number of sampling steps required. Additionally, a quicker version known as AVIS Flash is presented. This research tackles the challenges associated with zero-shot video inverse problems, aiming for effective real-time application.

Key facts

  • AVIS reduces initial latency from 114s to 4s.
  • AVIS increases throughput from 0.71 to 1.18 FPS.
  • AVIS uses autoregressive video diffusion models for streaming restoration.
  • AVIS initializes reverse diffusion with measurement-consistent estimate.
  • AVIS Flash is a highly accelerated variant.
  • The framework eliminates latency bottlenecks in video inverse problems.
  • AVIS achieves superior restoration quality compared to non-autoregressive solvers.
  • The method reduces required sampling steps.

Entities

Sources