AI upscale

While not a shader or even anything that we’ll see in our life time , I was mesmerized by these that I couldn’t help myself not to share…

on somewhat related note , is it possible to have bilinear shader with 1:1 color vibrance as the raw image ?

3 Likes

pretty wild stuff. Do you know what model/AI made them?

I’m not sure I follow. Can you elaborate?

This guy made them ; https://www.reddit.com/r/StableDiffusion/comments/13f7jy7/i_remaster_street_fighter_old_sprites_using/ according to him , The AI model tho called “controlnet 1.1” is made by this guy; https://twitter.com/TDS_95514874/status/1656637104749953026

In contrast to Scalefx/xbr , a lot of color is lost with bilinear . say , is it possible to make a shader version of your mednafen 3-point texture filtering for 2d games ?

1 Like

The result is very good. It is a dream to make textures, remapping.

I’m more surprised by the MK ones, although, I don’t know why the AI thinks humans have 5 fingers.

1 Like

Didn’t even notice that , can’t unsee it now I guess

I am not surprised, anomalous fingers and crooked eyes are common in AI.

image

The hands in the illustration is very good.

But I’m curious about Chun-Li’s original pixel-art, in which classic 2D game does he have his hands open?

He has the face, hair and colors of SFII.
The shirt and boots from SFIII.
Some parts of XvSF.
That white hair bun with no shadow.
The most is the position of the left foot.
But what visually shocks me the white belt and open hands like SFIV.

And that weird white border, as if they cropped the image wrong.

The image is AI-generated, but it is not upscale.You know what’s more likely. They took the SFIV 3d image, converted it to illustration and then to pixel-art.

Meme time:
image
"I don’t know Rick, it looks fake"

1 Like

As impressive as it might be, I can’t consider this an upscale. It’s more like a reimagine or remake than an upscale. It introduces new information not present in original. An upscale just fill the gaps, so it’ll never create a new finger from original image.

3 Likes

I agree with that. So called “AI upscaler” are replacing parts of the image with a new construct based on existing full images. Whereas an upscaler is based solely on the pixels of the original image, plus some math to form a new expression of it. No other images or parts of it is used to replace a face or eye in example.

I want to add another thing: AI is not predictable and may have different output with each request/calculation. A specific Shader algorithm based on the existing pixels only will always produce the same output.

1 Like

Don’t they just identify the pixels as “Chun Li” and then create a picture of Chun Li based on pictures it has found on the internet?

edit I guess that comment has already basically been made already.

1 Like

Who published the images says “I remaster Street Fighter old sprites”.

What I want to say is that it is false.
There is no such image in any classic 2D game (it’s not even in Mugen).

It was probably created by a pixel-art from the SFIV image.

1 Like

based on a quick google image search , I believe that particular Chun-Li pose , is fan made by this Artist :

https://ckelsallpixels.artstation.com/projects/KaJeoy

Other sprites origins are legit tho , I can vouch for kof mary

So it basically makes the graphics looks like shitty cheap phone-tier generic anime graphics that you would see on a shitty gacha. Burn this shit with fire.

2 Likes

It’s not these particular results but what could be done with AI in the far future , to me , kof13 has some of the best sprites I’ve ever seen , the same technique could be used to produce that type of sprites or anyway you wish . the mortal kombat looks nice imo.

There is debate to be had about “authentic look” if the same game could be redrawn in infinite amount of ways.

Than again , for this to happen you need gpu to apply esrgan 60fps without adding lag , I don’t see this happening. let’s stick to the crt’s for now.

This is the description I had in my soul, I can’t imagine all games with the same graphics.

It depends…

In Flycast games run completely smooth with 1080 textures in the integrated i9, with a GTX960 you can put more than 4K.

What I see with these images is very simple, it is totally false, there are five hundred videos scattered in the networks and no one says how he did it. And they are all the same.

1 Like

Not sure about that.

First, because one thing is seeing a still upscaled frame, another is to see the whole animation; I don’t know if used AI tools are able to track objects in motion; the risk is to see glitches near their borders when the AI is unable to distinguish between the sprite and the background.

So, the idea is to pre-scale single “sprites”, but then again, how they will look in movement is yet to be seen, right? What would ensure us that a bandana will not turn into hairs and back from frame to frame? look at Sonia.

My personal opinion, after an initial wow effect, is that this is the fastest way to kill pixelart.

1 Like

You’re talking about texture replacement where the assets are already upscaled and ready to be used and injected , there are some sprites replacement projects for nes mesen emu, most of which has flash games look

In contrast , real time AI upscaling during gameplay simply hasn’t been done yet . Dlss or fsr doesn’t count as the game is communicating vector information rather than upscaling entire frames.

Agreed , in blade runner future tho, perhaps the AI is smart enough to recognize the intended shapes that it upscale.

Like everyone else here , I think the retro games were made with crt in mind.

Imo long-term the best way to upscale pixel art is brute-force nearest-neighbour scaling on a 10000 to 20000PPI display, or maybe we would need an atomic resolution display where the pixels are as small as atoms

a square is a square, if it’s made of 4 pixels or 4 millions it will always be a square, unless you see it on a tiny display where our eyesight limit kicks in.

2 Likes

Except that’s not how it works. The smaller the pixels are and the smaller the gaps between the pixels become the smoother the image itself becomes(as well as the scaling which will have less and less noticeable jaggies and will start to look as smooth as upscaled square/straight line graphics when you have a high enough density). Yes, there is only so much detail with 240p and the sweet spot depends on viewing distance/screen siz4 combination, but NN upscaling to an extremely high PPI display is the best possible scenario long-term if you care about maintaining authenticity.