Adobe has just unveiled VideoGigaGAN, a new generative AI model capable of rendering blurry videos sharp and detailed. This impressive technology can increase the resolution of videos up to 8 times while avoiding the usual artifacts associated with super video resolution.
Adobe Research engineers have taken the next step in improving video quality using artificial intelligence. Indeed, they have just presented VideoGigaGAN, a revolutionary generative AI model capable of transforming blurry videos into sharp and detailed sequences, with a resolution up to 8 times higher than the original.
This major advancement combines the power of generative adversarial networks (GAN) and new techniques to preserve temporal coherence, avoiding the usual artifacts of video super-resolution. Although still at the experimental stage, VideoGigaGAN opens up promising prospects for video professionals.
A unique technology
VideoGigaGAN builds on GigaGAN, a large-scale image super-resolution model trained on billions of images. By combining the power of generative adversarial networks with new temporal coherence preservation techniques, VideoGigaGAN manages to generate ultra-detailed videos while avoiding common artifacts like flickering or distortion. Unlike previous video super-resolution approaches, which often produce blurry results, VideoGigaGAN achieves the feat of balancing sharpness and consistency.
The ingenuity of VideoGigaGAN lies in its unique architecture. The model integrates temporal attention layers into the decoding blocks to reinforce temporal coherence as well as an optical flow guided propagation module. An anti-aliasing block is used in the encoder to remove aliasing artifacts. Finally, a skip connection allows high-frequency characteristics to be transferred directly to the decoder layers, thereby compensating for the loss of detail. Thanks to these innovations, VideoGigaGAN pushes the limits of video super-resolution.
What uses for VideoGigaGAN?
Although, at this stage, VideoGigaGAN is still only a research project, its potential uses are numerous. For video professionals, this technology could significantly improve the quality of old sequences or those shot in difficult conditions. Integrated with editing software like Adobe Premiere Pro or After Effects, it would offer a quick and effective solution for “saving” low-resolution shots. VideoGigaGAN could also be used to restore old films, giving them a second life in high definition.
Beyond the professional world, VideoGigaGAN could also find its place among the general public. Indeed, who has never dreamed of being able to improve the quality of their old family videos and other travel memories? With the democratization of social networks, coupled with the craze for artificial intelligence, the demand for equally effective video editing tools continues to grow.
And if Adobe decides in the not-so-distant future to integrate VideoGigaGAN into its consumer applications like Premiere Elements or Photoshop, there is no doubt that this technology will be very successful. And while it should be kept in mind that using AI to manipulate visual reality and retouch people raises ethical questions that will need to be considered, other players such as NVIDIA, Microsoft and Blackmagic Design are also working on AI video super-resolution technologies, suggesting increased competition in this area.
Sources: Adobe, GitHub
6