[LOCKSUIT]

I beat the site's output! https://letsenhance.io/boost

SR=super resolution !

We should make a world record for this, and see who can top each other !!!!!


The first pic below is a very small bird. Pic 2 is what Let's Enhance can give me. But look at pic 3 - by shrinking it back down by some trick I got the neural network to think it was a non-SRed pic and add even more features to it! Pic 3 is better!! Pic 4 is even possibly better still. I tried for a day to keep it going but I give up, for now...

It will be fun to see who here can get the bird bigger. And in the future new algorithms can show the result here and look back at how big it got. So far Lock has the biggest bird.

[LOCKSUIT]

photo 4:

[LOCKSUIT]

I figured, if their SR can add some further details, and make it bigger/more 'better', then it can keep doing so !!. Btw, below is 3 SRs of theirs just straight 3 applications, yes it a bit better than 2 applications but mine still rules, and on 4 applications it actually gets exponentially less better and actually worse and goes watercolor blurry. Their SR CAN keep doing it, adding new details even after 20 applications, but, what's added is not desired to zoom into a bird, the desired zoom in slows and reverses. Also, the 2nd pic below is another as good if not better one, actually with less steps to reproduce it, and has same colors/lighting too! All I did to get this one is SR the original small bird, downsize it by half using bicubic, SR it, downsize it by bicubic like 10 times until get to size of 241 and on some/every other downsize maybe i did best quality and not bicubic this time, SR it a 3rd time, and lastly - send it through waifux2 LOL: http://waifu2x.udp.jp/

[LOCKSUIT]

2nd photo:

[LOCKSUIT]

Ohhh, it seems that vid of those growin celeb faces really can use a source image!?:
https://arstechnica.com/information-technology/2017/02/google-brain-super-resolution-zoom-enhance/

It uses pixelCNN? Where is the git implementation? Korr can setup it up? I sent someone a email on github, we'll see.

[Korrelan]

No… no…no…

As the editor rightly points out, you can’t get something from nothing (excluding quantum), its impossible… end of.

I really wish they wouldn’t write crap like this, it’s a load of bo**ocks, Arstech should really know better by now.

The first part, the conditioning network, tries to map the the 8×8 source image against other high resolution images.

What they are not telling us is that the ‘ground truth’ high resolution images where part of training set.

So the system is basically just picking high resolution details from the training data/ image that best suits the 8x8 pixelated image.

 

[LOCKSUIT]

I didn't tell you 1 thing - I know it can only do some (possibly fairly good) guesswork like features a b c d e etc probably come with the given input features, but, just the mere fact it can SR things tons way more than Let's Enhance (even if they are away from ground truth) still is what I want - I want to in a sense create masterpieces from small photos or upsize photos I have and have some fun / see what will happen. It's also a good tool if it helps a bit show you ex. what a sigh says.

If I feed it a face not in the training data, it can probably make my 40x40 face into a really bigger face, like the celeb ones that were amazing.