Introducing VideoGigaGAN, an innovative video super-resolution model capable of producing videos with high-frequency details and exceptional temporal consistency. Built on the powerful GigaGAN technology, our model surpasses traditional limitations, offering unparalleled video clarity and smoothness. Experience the future of video super-resolution today.
In the realm of video super-resolution (VSR), existing methods successfully maintain good temporal consistency in enhanced videos, yet often produce images that are blurrier compared to image upscaling techniques, primarily due to their limited generative capabilities.
This leads to an important question: Is it possible to apply the success of image generation upscaling to VSR tasks while also preserving video temporal consistency? To address this, we introduce VideoGigaGAN, an innovative generative VSR model designed to generate videos that are rich in high-frequency details and exhibit temporal consistency.
VideoGigaGAN evolves from a large-scale image upscaling model, GigaGAN, with significant enhancements. We discovered that simply adding temporal modules to extend GigaGAN to video applications results in severe temporal flickering. Addressing this, we identified and tackled several key issues, introducing a series of techniques that significantly enhance the temporal consistency of upscaled videos.
Through rigorous testing, VideoGigaGAN not only surpasses previous VSR technologies in maintaining video temporal consistency but also produces videos with finer and more detailed appearances. By comparing VideoGigaGAN with the latest state-of-the-art VSR models on public datasets and showcasing video results at an 8x super-resolution, we demonstrate the model's superior performance.
