Generative video compression just hit 60 FPS on 1080p, slashing bitrates by 85% without the typical diffusion 'lag.'
April 16, 2026
Original Paper
CoD-Lite: Real-Time Diffusion-Based Generative Image Compression
arXiv · 2604.12525
The Takeaway
Diffusion models are usually too slow for real-time video, but CoD-Lite achieves a 'one-step' codec that runs at 60 FPS for encoding. This allows for massive 85% bitrate reductions compared to current SOTA generative codecs, while finally being fast enough for live streaming. This is a huge leap for practical AI in telecommunications. It means we can stream high-definition video over incredibly weak connections by having the AI 're-imagine' the details on the other end in real-time. This transforms how we think about bandwidth in remote areas or high-congestion environments. It turns a theoretical AI advantage into a production-ready tool.
From the abstract
Recent advanced diffusion methods typically derive strong generative priors by scaling diffusion transformers. However, scaling fails to generalize when adapted for real-time compression scenarios that demand lightweight models. In this paper, we explore the design of real-time and lightweight diffusion codecs by addressing two pivotal questions. First, does diffusion pre-training benefit lightweight diffusion codecs? Through systematic analysis, we find that generation-oriented pre-training is