/sci/ - Science & Math

>>10247951

Could you half-life grunts? Wonder what details it would bring out

zoom in in next image post of the bottom left

>>10247961

last image post unless people want to see more. The advancement of AI into best-in-class upscaling was pretty rapid.

>>10247970
More fgt

>>10247990

>>10247994
Can I use this to play elder scrolls 2 with good graphics?
Anyone try Doom?

>>10247994
>>10247970
>>10247966
>>10247961
>>10247952
>>10247951

impressive... very nice.

>>10247951
>zoom and enhance is about to become a reality
wtf

>>10247961
>>10247966
Zoom and enhance!
https://www.youtube.com/watch?v=gF_qQYrCcns

>>10247994
I could fap to that.

>>10248016
>>10248098
>not enhancing noise into optimized fapping material

Is there somewhere you can do this online?

>>10248142
Soooooon

>>10248016
your brain is literally zoom&enhancing constantly.

Waifu2x

The biggest hurdle was probably collecting and cleaning all the training data.

>>10248271
https://topazlabs.com/

>>10248349
More than 90% of their products are things you can already do in photoshop in literal minutes
Why would anyone buy a black and white plugin?

>>10248359
they use this
https://topazlabs.com/trials

Shit's cool Anon. Post more

>>10247951
Fake and gay

I want to believe.

>>10248397
This
It's 100% fake

I don't know if I'm more impressed by the upscaler or the pixel artists who were able to create art with enough information that these images could be generated from them.

>>10248349
>>10248361
>No Linux version.

Is there any documentation about what research it's basing it's approach on? Maybe there is another implementation of the technique on github somewhere.

>>10248088
lol, reminded me of this
https://www.youtube.com/watch?v=3uoM5kfZIQ0

>>10248397

It takes several seconds to generate a picture and thus isn't viable for gaming yet

>>10248496
Oh also there's no continuity between two generated frames either, so you'd essentially see a fuckton of swirling in the objects.

>>10248496
>It takes several seconds to generate a picture and thus isn't viable for gaming yet
No shit. It's probably a GAN. Only retards on /g/ and /v/ blurt out retarded shit about gaming as soon as they see stuff like this.

>>10248499
There are GANs that work over sequences of frames and I'm pretty sure there should be no problem generalizing the technique here. Though I don't know why you'd want that desu.

>>10248496
Could easily make an automatic texture pack generator though. But with all the single pics from this one and ESRGAN I havent seen anyone do a full pack for any game...

>>10249011
morrowind
https://www.nexusmods.com/morrowind/mods/46221?tab=description

Gothic 1 https://www.reddit.com/r/GameUpscale/comments/a9ssit/gothic_1_esrgan_project/

a doom 1 version
https://www.doomworld.com/forum/topic/99021-v-0-95-doom-neural-upscale-2x/

>>10249036
Ah ok I guess it still doesnt look like much since the models are still low polygon count. Anyone know how hard it would be to use this kind of machine learning to work with 3D models? It should kind of be the same principle although probably much more time consuming...

>>10249070
depends whether they were modeled low-poly to begin with

then again i'm just assuming this upscaling method relies on the images having once been high resolution and then downscaled, leaving behind contributions that can be identified by the network. maybe that's not how it be.

>>10249133
Nah thats not how it works as far as I know. I read through the paper behind esrgan a bit and it basically looks for objects it recognizes and adds features to that area that are in line with the object.
They have some examples on their github page where they compare original HD images with downscaled and then put through esrgan versions. One of them was some form of mansion with open windows in the original which werent really noticeable at all anymore in the downscaled version, but then reappeared with closed windowbars after esrgan upscaled it again...
Also I doubt these old graphics are downscaled high resolution images. Most likely they were drawn pixel by pixel like with modern pixel art.

>>10247951
Any Morrowind?

>>10247951
how does this compare to waifu2x? Or are they made for separate use cases?

I'd love to see an AI-based automatic tris-to-quads converter and/or "sharp edges finder" for constraining Catmull-Clark subdivisions as a similar process to this in 3D.

>>10250569
There’s an entire texture pack available for morrowind.

>>10247966
If I see that fucking bottle one more time,

>>10250804

>>10252046

>>10247970
This is bullshit. I can tell because, for it to reconstruct the ribs like that, it would have to know that its recomstructing imagery of a ribcage. For a system to have insight like that it would have to have pretty good quality wide generalization and reasoning, basically would be full AGI.

So either the neural net that makes these images is a few steps from being a full blown AGI, or this is bullshit and either a higher resolution skelly than we see there was used, or it was run 1000 times on the skelly and they chose one where it happened to look really good.

I dont think we are nearly that close to agi so its probably bullshit

So it's 5:30am and I finally got this shit working
It's not Topaz like OP said- it's called ESRGAN and it's a Deep Learning tool

Again, I'm tired as fuck so I'm sure I did something wrong/haven't read the documentation closely enough, but I haven't been getting fantastic results

Went ahead and made a quick comparison

Here's my comparison of all the different methods mentioned ITT

>>10254035
Topaz wins, but only since it adds an artistic touch to it. The original is better otherwise

>>10254041
for my money, in terms of setup time and price, waifu2x is still the best option
Topaz is objectively the winner, but it's not $100 better

>>10254044
I'd keep the original desu, it looks fine

>>10254058
Agreed
I just meant in terms of upscaling

>>10252050
>>10252046
Wow, it's fucking nothing

>>10254141
t. spoiled baby

>>10254035
Decided to run a comparison using a real photo
ESRGAN seems geared toward this type of work, but TOPAZ about ties it up

It's interesting to see how the two different methods attack the same problem

>>10254035
>>10254740
So does this mean OP's posts are bullshit?

>>10248457

it's almost certainly a CNN set up in a GAN configuration. i actually thought about doing something like this a while back, since you don't need an annotated dataset, just a big set of images that you can downsample/upsample for input and then train with the originals as the targets.

>>10254777
There are some bullshit images going around itt.

>>10254740
Kek

>>10254035
the second one kinda looks cute

Brainlet here, how long before technology like this can be used to cockblock game developers who try to sell HD remasters by making free fanmade texture packs for every game?

>>10255859
Interesting point. I'd love to see this used on Phoenix Wright or Professor Layton and see how it squares up to the official HD remasters

>>10256007
Nintendo would never allow it.

>>10256183
not officially you mean

>>10257264
>>10256183
I only meant comparing one or two screencaps, not redoing the entire games

>>10248499
nah, adding optical flow to GANs is fairly simple nowadays

>>10248496
So what's stopping you from generating them beforehand?

Possible evil appliances: Lying about graphics in pre order promo video.

Evilest AI that actually exist.

>>10248496

the idea is that you could upscale textures/sprites in order to replace the lower quality original versions.

>>10247951
Can you try it on this?

It's interesting how AI handles this problem.

>>10260015
400% upscale (150-200% is ideal)

Hear me out-imagine a lossy compression method using this to enlarge images that have been shrunk down to decrease size.
thoughts?

>>10260521
https://arxiv.org/abs/1609.04802

https://arxiv.org/abs/1809.00219

Damn this actually delivers

>>10260486
Is that Taeyeon?

Try it on censored porn.

>>10260494
Beautiful.

>>10247970
I want a general for this