Accelerating Video Inverse Problem Solvers with Autoregressive Diffusion Models
The paper introduces the Autoregressive Video Inverse problem Solver (AVIS), which aims to enhance the efficiency of video restoration using autoregressive diffusion models. AVIS significantly reduces initial latency and increases throughput compared to existing methods, making it suitable for real-time applications. An accelerated variant, AVIS Flash, further improves performance while maintaining quality, paving the way for practical deployment.
- ▪AVIS reduces initial latency from 114 seconds to 4 seconds.
- ▪Throughput increases from 0.71 to 1.18 frames per second (FPS) with AVIS.
- ▪AVIS Flash boosts throughput to 5.91 FPS on a single RTX 4090 GPU.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.20624 (cs) [Submitted on 20 May 2026] Title:Accelerating Video Inverse Problem Solvers with Autoregressive Diffusion Models Authors:Taesung Kwon, Jonghyun Park, Hyungjin Chung, Jong Chul Ye View a PDF of the paper titled Accelerating Video Inverse Problem Solvers with Autoregressive Diffusion Models, by Taesung Kwon and 3 other authors View PDF HTML (experimental) Abstract:Diffusion models provide powerful priors for zero-shot video inverse problems, but their real-time deployment is hindered by two inefficiencies: high initial latency caused by holistic video restoration, and low throughput resulting from multiple VAE passes to enforce measurement consistency in pixel space.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
