Nvidia: AI Reconstructs Photos with Realistic Results

Published by

Click here to post a comment for Nvidia: AI Reconstructs Photos with Realistic Results on our message forum
https://forums.guru3d.com/data/avatars/m/258/258664.jpg
I mean, this looks good on the small scale, but I don't think they're yet there to have it done in a quality that couldn't be identified as "fake"
data/avatar/default/avatar10.webp
fantaskarsef:

I mean, this looks good on the small scale, but I don't think they're yet there to have it done in a quality that couldn't be identified as "fake"
Would agree with you there, i think humans still do better than this for the time being.
data/avatar/default/avatar22.webp
It depends what is the goal and how much resource you can spend. Human can do better, of course, but takes much more time or costs more. e.g. removing watermarks or an electric wire / pole from a beautyful panorama view a removing a funny photobomber takes only a few seconds with this tool, but manually can take hours for an amateur and even long minutes for a pro. And for an amateur or facebook/instagram user time is more important than quality.
https://forums.guru3d.com/data/avatars/m/248/248994.jpg
That was decently impressive, but I'd say only assuming one could do that at home with their regular Nvidia video card and without needing to pay hundreds, if not thousands, of dollars for the software. Of course it being a little slower would be acceptable, depending on the GPU. If the software was published only to professionals working for big companies, like media houses, and they needed to pay big money to use it for a year... Nah, I wouldn't call it impressive anymore.
https://forums.guru3d.com/data/avatars/m/269/269560.jpg
Whaooouhh. I definitely know now who to contact to get a younger and nicer picture of me, myself and I for my LinkedIn profile.
https://forums.guru3d.com/data/avatars/m/243/243702.jpg
Pawel04:

Would agree with you there, i think humans still do better than this for the time being.
Better, but nowhere near as fast.
https://forums.guru3d.com/data/avatars/m/80/80129.jpg
lucasfrance:

Check https://www.theinpaint.com/ the Inpaint sw does this for a long time alsready !
Inpaint and tools like content aware fill in photoshop sample the surrounding area and just paste it over the target you're trying to remove. It works for most things but you can see the patterns it's sampling from and won't work in huge areas. The technique Nvidia is using actually generates imagery based on what's being targeted from a library of over 20K pictures. If you used inpaint to replace the eyes on a person, or photoshop, it wouldn't generate new eyes like this does. It's a similar output but how the technique is being performed is way different. With a larger dataset and more training the Nvidia technique will be far more accurate and less prone to 'patterning"
https://forums.guru3d.com/data/avatars/m/270/270492.jpg
Denial:

Inpaint and tools like content aware fill in photoshop sample the surrounding area and just paste it over the target you're trying to remove. It works for most things but you can see the patterns it's sampling from and won't work in huge areas. The technique Nvidia is using actually generates imagery based on what's being targeted from a library of over 20K pictures. If you used inpaint to replace the eyes on a person, or photoshop, it wouldn't generate new eyes like this does. It's a similar output but how the technique is being performed is way different. With a larger dataset and more training the Nvidia technique will be far more accurate and less prone to 'patterning"
You're right! Inpaint is clearly not able to create "eyes" as in the Nvidia demo... But it works well for a picture small area cleaning.
https://forums.guru3d.com/data/avatars/m/63/63215.jpg
I like it. For consumers, the quality is acceptable, especially for most of the scenery stuff. On humans, it's not as effective...yet.
https://forums.guru3d.com/data/avatars/m/246/246171.jpg
I personally think this is really cool. As others pointed out, this isn't just simply making due with whatever textures are currently available in the photo. Though I found the eye replacements a little bit shocking (even a bit funny), I also found it to be pretty fascinating. Furthermore, this was being run on a Linux system, which I personally like since (to my knowledge) there is no Linux-native equivalent to this kind of thing. But seeing as this is Nvidia, I can't help but think this is going to use CUDA or some other proprietary tech. I would think this would mostly be a CPU-intensive task, or at least not all that beneficial of GPU use (regardless of brand).
https://forums.guru3d.com/data/avatars/m/242/242471.jpg
This Iooks perfect for my scenery editing pics 😀 Where can I download this??
data/avatar/default/avatar34.webp
Great for stills. Call me when it works on VIDEO, which is where the real challenges are in this space.
https://forums.guru3d.com/data/avatars/m/68/68055.jpg
Content aware stuff in Photoshop worked always perfect in Adobe presentations too, but in real life sometimes it acts stupid, and it's much easier to stamp it manually.
https://forums.guru3d.com/data/avatars/m/216/216490.jpg
What sorcery is that! Now the future AI will even be able to erase all digital footprint of ours! *oops had my tinfoil hat on again. I swear this thing jumps on me head by its own! P.S: Cool tech and video btw. Props.
https://forums.guru3d.com/data/avatars/m/165/165326.jpg
Great to see this AI technology advancing. maybe in 5 ~10 years it will be perfect. ( i ROFL at the guy with the huge hairy eyebrows , the AI trimmed it for him 😀 )