The app boosts video quality up to eight times without annoying AI noise.
A team of video and artificial intelligence engineers from Adobe Research has developed an artificial intelligence application called VideoGigaGAN. The device can take a blurry video sample and, after processing, return the same sample with increased sharpness and clarity, reports TechXplore.
The team used a generative adversarial neural network (GAN) to teach the system what clear, sharp video looks like (for example, an individual eyebrow hair rather than a blurry mass). Then, a “flow-controlled rotation module” was added to ensure consistency between video frames.
Generative adversarial networks are composed of two neural networks: a generator and a discriminator, which compete with each other during the training process. The generator tries to “fool” the discriminator by producing data that the person mistakes for real data. In doing so, the discriminator tries to be more accurate in differentiating between real data and generated data.
The scientists also used anti-aliasing techniques to prevent AI glitches and “high-frequency detail redirection” to deal with sudden drops in video quality.
According to the researchers, the system can increase video quality up to eight times. And all this without unnatural colors, sharp lines, and other common problems encountered when processing video using artificial intelligence.
The developers acknowledge that part of the output video is fully synthesized by VideoGigaGAN based on system ratings. For example, pores on the skin, wrinkles around the eyes, or even eyes can be added to the video to achieve high definition. That is, the AI ”completes” the necessary elements.
It is unclear whether Adobe will release the app for general use.
Previously Adobe presented neural network for creating music. Project Music GenAI Control allows you to generate audio based on a text description.
