Create Super Resolution Video Contents using Generative AI

Standard approaches for upscaling video content involve complex processes, including frame interpolation, denoising, filtering, post-processing, and others. Two notable approaches for upscaling video resolutions are: 1) Using COTS (Commercial Off-The-Shelf) licensed products; 2) Building a video upscaling pipeline using open-source tools like FFmpeg. Using COTS products could be costly due to licensing costs. Additionally, it presents major challenges when a large number of media contents require video upsampling. Despite COTS products offering streamlined approaches for video upscaling tasks, these products are built using proprietary technology and vendor lock-in. On the other hand, leveraging open-source tools could be an alternative that provides better flexibility and control, but the results could be inconsistent. Recent advancements in Deep Learning and Generative AI technology have made super-resolution images through AI possible. These models are trained with millions of images that learn common patterns for high resolution. These models are ideal for achieving upscaled image quality at scale through a single pass without explicit steps for denoising, filtering, and so on. Generative AI techniques, such as Generative Adversarial Networks (GANs), Vision Transformers, and Stable Diffusion, are considered the front runners when it comes to using AI techniques for image super-resolution tasks. In this demo, we showcase an end-to-end solution that uses Real-ESRGAN, SwinIR, and is coupled with AWS services to orchestrate a workflow that takes low-resolution videos as input and produces 4x resolution videos in minutes. The solution could be used to automate the process of super resolution for media contents at scale.