Adobe researchers have developed a brand new generative AI mannequin known as VideoGigaGAN that may upscale blurry movies at as much as eight instances their unique decision. Launched in a paper published on April 18th, Adobe claims VideoGigaGAN is superior to different Video Tremendous Decision (VSR) strategies as it will possibly present extra fine-grained particulars with out introducing any “AI weirdness” to the footage.
In a nutshell, Generative Adversarial Networks (GANs) are efficient for upscaling nonetheless pictures to a better decision, however wrestle to do the identical for video with out introducing flickering and different undesirable artifacts. Different upscaling strategies can keep away from this, however the outcomes aren’t as sharp or detailed. VideoGigaGAN goals to supply one of the best of each worlds — the upper picture/video high quality of GAN fashions, with fewer flickering or distortion points throughout output frames. The corporate has supplied several examples here that present its work in full decision.
A few of the finer particulars within the demo clips Adobe supplied look like fully synthetic, such because the pores and skin texture and creases within the under instance, however the outcomes seem impressively pure. It will be tough to inform that generative AI was used to enhance the decision, which might prolong the “what is a photo” debate to incorporate video.
That is solely a analysis preview, so there’s no assure that Adobe will make VideoGigaGAN obtainable to shoppers by way of Inventive Cloud software program like Premiere Professional. The corporate beforehand previewed a separate diffusion-based upsampling experiment, Project Res-Up, throughout its MAX occasion in October 2023, which equally improves the standard of low-resolution GIFs and video footage. And Adobe isn’t alone on this work, as each Microsoft and Nvidia have additionally developed their very own VSR upscaling know-how.