Gigapixel AI - better face and hair restoration/enhancement

It would be great if Gigapixel would be capable of doing some better face enhancement.

I’m a hobbyist retoucher and an upscale addict. I discovered Gigapixel while I was looking for a tool that would help me increase the quality of portrait photos of my late friend. That’s how my adventure with photo editing began. I started joining Facebook groups related to photo editing, and today in my free time I tend to process dozens of pictures every day.

In my work, I very rarely have occasions to work with professional images. In many cases, I deal with photos taken with very low quality and cheap smartphones and cameras. Of course, software such as Gigapixel, Sharpen AI, or Denoise AI can do miracles even with extremely bad photos, however, especially in portrait photos, there are always some elements that don’t let me go to bed and fall asleep peacefully at the end of the day, namely: faces and hair.

So, that’s why I keep on looking for a tool that would enhance faces and hair in upscaled photos. I found some tools like Remini (an online service with a client app for smartphones, they also provide some browser-based version ‘for professionals’), DFDNet, or PSFRGAN (both can be found on Github). They all perform a perfect job on faces, but the problem is that they only process face and some square area around it, leaving most of the hair untouched and looking like the subject put a ton of glue on their head. What’s more, DFDNet and PSFRGAN are extremely limited (they can only process faces that aren’t bigger than 512x512x), and Remini is horrendously expensive (in the browser-based version of their service, processing SINGLE image costs 1$) or uncomfortable to use (you need a smartphone or an Android emulator to use it on PC).

That’s why I hope that Gigapixel will receive some better face and hair enhancement algorithms at some point in the future.

Picking up on this old post;

I too would like to see some progress with hair and skin texture.

I think actually hair is more difficult as the problem is that upscaling makes hair strands thicker and is unable to “invent” new strands. The subject’s hair turns into strands of string! On the other hand, if it invented new strands, it might go horribly wrong. Nevertheless it would be good to have something, perhaps even experimental here, as even if it went wrong elsewhere, two images could be blended.

For skin texture, this is another one. I think this is a bit easier. Could just be more training on such things. Even if it totally invented new skin texture, it would most likely be an improvement. As it is, upscaling skin gives a “plastic” look.

I’ve also tried “Remini” and it does get some good results. Obviously trained mostly on portraits. Which of course, would be a very useful mode!

1 Like