NVIDIA AI Fixing Grainy Pictures

FrgMstr

FrgMstr

Just Plain Mean
Staff member
Joined
May 18, 1997
Messages
55,634
While some of us have some incredibly complex smartphones that take beautiful pictures, there are still those times when lighting causes us to get less than favorable shots. While this tech is not available to us yet, and likely will not be for a while, the folks at Aalto University, MIT, and NVIDIA have been working on a solution to getting those photos touched up without even having to see "clean" ones for comparison. That said, the technology was trained by looking only at corrupted photographs, which is a bit wild when you think about it. While everyone loves a great photo, these techniques are also looking to be used in the medical field, but of course in those instances it has to be exactly right.

Check out the video.


“There are several real-world situations where obtaining clean training data is difficult: low-light photography (e.g., astronomical imaging), physically-based rendering, and magnetic resonance imaging,” the team said. “Our proof-of-concept demonstrations point the way to significant potential benefits in these applications by removing the need for potentially strenuous collection of clean data. Of course, there is no free lunch – we cannot learn to pick up features that are not there in the input data – but this applies equally to training with clean targets.”
 
Some of those are so impressive It's hard to believe. I wonder if the results are typical or cherry picked and how much training is required - would it only work in a cloud scenario or could your phone do it.
 
andrewaggb said:
Some of those are so impressive It's hard to believe. I wonder if the results are typical or cherry picked and how much training is required - would it only work in a cloud scenario or could your phone do it.
Click to expand...
Of course they're cherry picked, what do you think? Promo results are always cherry picked no matter the subject matter. The question is how good the average result is. Which nobody will know until they release the technology. Oh, it is nvidia they won't release it to the public.
 
Although most of those are showing impressive results. The overall is impressive but I suspect it will be far from professional quality photos. Too much interpretation is needed to fill in the removed noise and add in the correct color and sharpness.
 
The MRI one will never happen, they will never allow "best guess AI data back fill" into medical data. Sure the picture looks better and more pleasing, but it's not necessarily accurate after that AI interpretation.
 
xorbe said:
The MRI one will never happen, they will never allow "best guess AI data back fill" into medical data. Sure the picture looks better and more pleasing, but it's not necessarily accurate after that AI interpretation.
Click to expand...
I don't think they'd suggest using a corrected image in lieu of the noisy source, but having access to both could certainly be beneficial.
 
Looks to me like the headline should be "Nvidia AI learns to compensate for Nvidia's noise-adding algorithm" :D
 
I keep seeing Nvidia headlines similar to this which other software already have accomplished. Kudos to them though for creating competition but the fact their software\hardware can accomplish this is nothing new.
 
I can see this being abused to remove water marks from pictures taken by the actual author of said content.
 
I wish Nvidia would release the source (or at least the compiled versions) of all these demos they do. I'd love to try them out myself.
 
andrewaggb said:
Some of those are so impressive It's hard to believe. I wonder if the results are typical or cherry picked and how much training is required - would it only work in a cloud scenario or could your phone do it.
Click to expand...

Yeah, that's definitely way better than traditional denoise filters, which typically just amount to a light blur.

They are almost too good to be true...
 
SixFootDuo said:
Shrug .... I haven't seen grainy photos since 2005
Click to expand...

Try taking pictures in very low light. Concert photography, etc. Granted, the latest very high end DSLR's have remarkably good low noise performance at high ISO, but they are also ridiculously expensive. Even a mid range pro-sumer DSLR, not to mention a cellphone camera is either going to have motion blur or high ISO noise when it gets darker.
 
Zarathustra[H] said:
Yeah, that's definitely way better than traditional denoise filters, which typically just amount to a light blur.

They are almost too good to be true...
Click to expand...

Yeah, those demo results are really good, but sharpening after using a good "traditional" denoiser isn't too bad.

Waifu2X is also a good stand-alone denoiser, but it's neural-net based like this Nvidia demo.

http://waifu2x.udp.jp
 
andrewaggb said:
Some of those are so impressive It's hard to believe. I wonder if the results are typical or cherry picked and how much training is required - would it only work in a cloud scenario or could your phone do it.
Click to expand...

Some of the results are also a dumpster fire of oversmoothing. About the only thing I can say about both the good and bad examples is that they appear to largely avoid inducing sharpening artifacts. Which is a plus.
 
Pretty cool but I wonder has this AI contacted Ted Turner to further his touch up projects?

 
The corrupted example was pretty impressive. So those with no additional data huh? Nice.
 
YUjACfF.gif
 
CSI Miami is becoming a reality!

"Computer..... Enhance resolution."

"Well, I guess we ... resolved this case.... " [puts on sunglasses]

[Cue the music]
 
anything from an iphone 6 and up have excellent photos. The 5 still takes cheese a lot at distance.
 
You must log in or register to reply here.
Back
Top