How ESRGAN fits into restoration?
I am experimenting with the Real-ESRGAN (by Xintao Wang), which stands for Enhanced Super-Resolution Generative Adversarial Network. This is how the tech is described in a paper:
Quote:
I have been experimenting with it within a dev copy of Selur Hybrid (available upon request my IM at Selur's discussion forum). My focus is on 8mm home video tapes, and some VHS second generations of am 8mm tape. I've tried ESRGAN (which has different models inside to choose from). The one I liked best is realesr-general. There is also another Resizer option in Hybrid called VSGAN, Single Image Super-Resolution Generative Adversarial Network (GAN) which "uses the VapourSynth processing framework to handle input and output image data." I tried a bunch of the models but I think the ESGAN realesr-general was best. I also compared ESRGAN to Topaz Video AI. Perhaps that is not a fair comparison, because Topaz Video AI doesn't only upscale but does the denoise, sharpening, dehalo and more. But I like to see what Topaz does to my 8mm video (which I deinterlace in Hybrid and upscale to 1080p before importing to Topaz, so that I can get access to Topaz enhanced tools like Proteus and Artemis that aren't available if the video is interlaced). After fiddling with a bunch of settings in Hybrid, I still can exactly replicate what I see in Topaz, which has more to do with my knowledge level on VapourSynth tools than anything with the power of those tools. I guess what I'm saying is that I find myself tempted to buy Topaz because I feel like I will never be able to figure out how to do it is well. I guess the dream is that one day something much better than Topaz will scan the video and apply the same knowledge and skill that the video restorer experts do and output a truly restored and enhanced copy of 8mm SD footage. If that is the case, I will keep my raw captures safe waiting for that day. |
Site design, images and content © 2002-2024 The Digital FAQ, www.digitalFAQ.com
Forum Software by vBulletin · Copyright © 2024 Jelsoft Enterprises Ltd.