Adobe researchers have developed a new generative AI model called VideoGigaGAN that can upscale blurry videos up to eight times their original resolution. Introduced in a paper published on April 18, Adobe says that VideoGigaGAN can provide more fine-grained details without introducing “AI weirdness” to the footage, compared to other video super resolution (VSR) methods. claims to be better than.
In a nutshell, generative adversarial networks (GANs) are effective at upscaling still images to high resolutions, but cannot do the same for videos without introducing flickering or other unwanted artifacts. is difficult to do. Other upscaling methods can get around this, but the results aren't as sharp or detailed. VideoGigaGAN aims to provide the best of both worlds, with higher image/video quality of GAN models and reduced flickering and distortion issues across the output frame. The company provides several examples of its full-resolution efforts here.
Although some of the details in the demo clip provided by Adobe look completely artificial (such as the skin texture and creases in the example below), the results look surprisingly natural. It's hard to tell if generative AI was used to improve resolution, and the “what is photography” debate could extend to video.
Since this is just a research preview, there is no guarantee that Adobe will make VideoGigaGAN available to consumers via Creative Cloud software such as Premiere Pro. The company previously previewed another diffusion-based upsampling experiment, Project Res-Up, at its October 2023 MAX event, which will similarly improve the quality of low-resolution GIF and video footage. Adobe is not alone in this effort, as Microsoft and Nvidia are also developing their own VSR upscaling technology.