Introducing Regen Video Upscaler

February 12, 2025 - Regen Research

We present Regen Video Upscaler, a method capable of enhancing and upscaling videos 4x their original resolution up to 8K, recovering fine-grained details with exceptional temporal consistency. On the challenging Panda-70M test set, Regen achieves better FVD/LPIPS metrics than any other available methods, marking a new state of the art for video super resolution. We confirm that these quantitative results are aligned with human preferences in a user study of 40+ industry experts, who showed an overwhelming preference for Regen results over the other methods, including the best commercial alternatives.

A general framework for video enhancement

The vast majority of videos ever produced was created in standard or HD resolution.
While modern AI video generation has made incredible strides, even cutting-edge models like Sora, Runway, Hailuo or Mochi can't exceed 1080p at 30fps, due to their computational constraints.
This leaves a vast and growing inventory of content unprepared for the demands of high resolution displays like 4K, 8K, and immersive AR/VR.
Regen provides a generalized solution to fill this technological gap, enhancing videos up to 8K with unprecedented detail recovery - from skin textures to fabric patterns that were previously invisible.

Method overview

Our pipeline is adapted from VideoGigaGAN (Xu et al, 2024), using a custom flow-guided feature warping pipeline inspired by BasicVSR++ (Chan et all, 2022). We trained our pipeline end-to-end on a custom large HQ videos dataset, using a series of advanced synthetic degradations extended from Real-ESRGAN (Wang et all, 2021).

Overview of the Regen Video Upscaler architecture

Quantitative evaluation

On a subset of 1,000 videos from the diverse and challenging Panda-70M test set, Regen achieves better FVD and LPIPS scores than the 4 other popular methods evaluated on the 4x upscaling task (BI), including 2 recent diffusion-based approaches that require significantly longer inference times.

FVD vs Inference time of Regen, BasicVSR++, Realviformer ,Upscale-A-Videoand StableVSR

User study

We confirmed that these improvements are aligned with human preferences in a user survey of 40+ industry experts who contributed more than 2.5k comparisons. The results show that Regen is overwhelmingly preferred over the other methods, achieving a global pairwise win rate of 93.5%.

The survey is still active and we welcome new contributions as well as new upscaling methods: Video Super-Resolution Arena

Win rate from 2.5k+ pairwise comparisons rated by 42 industry experts

Regen vs Topaz Video AI

Topaz Video AI is a popular commercial video upscaler, plebiscited by many professionals in the video editing industry. We conducted a qualitative comparison between Regen and various models of the Topaz Video AI solution, and found that Regen is able to recover more details with the same level of temporal consistency.

Comparison on 4x upscaling (480×270 → 1920×1080). Video by Jayanth Muppaneni, Pexels

In terms of user experience, Regen also has the advantage of being able to run directly from the browser, without the need to install any software.

Moreover, the Regen API allows direct integration of the Regen Video Upscaler with most video editing software and workflows.

Available today on Web and API

Regen Video Upscaler is available today via API and Web UI.

V2 is already under active development, with more customisation options and even sharper details.

If you'd like to work with us on custom integrations, let us know!

Make sure to follow @regenvideo on X for updates, there's a lot more coming soon.