Most of the new features and experiments that Adobe has recently announced are related to AI, such as adding and removing objects in Premiere Pro and text-based image generation in Photoshop. This time, the company announced his VideoGigaGAN, an experimental AI feature that can upscale videos by 8x without introducing the usual artifacts such as flickering and distortion. The Verge report.
According to Adobe, VideoGigaGAN is superior to other video super resolution (VSR) methods because it avoids the usual artifacts and flickering caused by General Adversarial Networks (GANs). At the same time, it adds clarity and detail as most other systems cannot do both at the same time.
Of course, this system is not suitable for things like forensic video enhancement, as it completely makes up details that don't exist. Ara CSI style crime show. However, the added details, such as skin texture, fine hair, and detail in the swan's feathers, look surprisingly realistic.
According to Adobe researchers, the model is built on a large-scale image upsampler called GigaGAN. His previous VSR models had difficulty producing rich details in the results. To that end, Adobe has introduced “temporal attention” (reducing artifacts that accumulate over time), feature propagation (adding detail to non-existent parts), anti-aliasing, and what it calls “HF shuttle” (shuttle). I combined it. high frequency functions) to create the final result.
Added to products like Premiere Pro and After Effects, video producers could potentially make low-resolution shots look much better, but using AI to emphasize people remains controversial there is. There's no word yet on whether Adobe plans to do this, but a number of companies (including NVIDIA, Microsoft, and Blackmagic Design) are also working on upscalers.
