– Adobe researchers unveiled VideoGigaGAN, a generative AI model that can upscale blurry videos into crisp, smooth video up to 8x sharper.
– VideoGigaGAN combines GigaGAN with temporal convolutional and attention layers to produce temporally consistent video with high-frequency details.
– The model separates low-frequency and high-frequency elements in each frame to remove noise and flickering, allowing for impressive results in upscaling video resolution.
Adobe researchers have developed VideoGigaGAN, an AI model that can upscale blurry videos to a much sharper quality, up to 8x. While image upscaling has been successful in the past, improving video upscaling poses a greater challenge due to the complexities involved in maintaining smooth transitions and high-frequency details. Current Video Super Resolution (VSR) techniques often struggle with balancing these aspects, resulting in videos that are either blurry or glitchy. Adobe’s VideoGigaGAN aims to overcome this issue by producing videos that are both temporally consistent and detailed.
The method relies on GigaGAN, a powerful generative adversarial network known for image super-resolution. Initially, applying GigaGAN to each frame of a video appeared promising but resulted in flickering and temporal inconsistencies. To address these issues, Adobe’s researchers incorporated temporal convolutional and attention layers into the network, as well as separating low-frequency and high-frequency elements within each frame. By smoothing out high-frequency details that cause noise and flickering in low-frequency features and preserving finer details through Skip connections, VideoGigaGAN effectively enhances video quality while maintaining consistency across frames.
Adobe’s GitHub demos showcase the impressive capabilities of VideoGigaGAN, which has not yet been given a release date. The potential applications of this technology are vast, ranging from historical archival footage to upscaling classic movies and TV shows. With further development and refinement, VideoGigaGAN has the potential to revolutionize the way we enhance video quality and preserve visual content.