Thread 106637777

Has image compression finally been solved?

>>106637786
is it supposed to look worse?

two minute papers is a jeet scammer

>>106637796
It looks like one of those shitty SNES filters.

>>106637796
Colors look more accurate but it also looks smoothed over. Might be good for some applications.

>>106637786
>brand new codec
>look inside
>gaussian blur

>>106637796
What exactly do you think the image you're looking at is?

>>106637817
His name is literally Károly Zsolnai-Fehér

>>106637777 (OP)
That's obviously a comparison against an uncompressed image. Now let's see how that fares againt a proper jpeg.

>>106637858
looks like its meant to be impasto oil painting, in which case the artifacts helps it give character vs the smooth detail-less image on the right. my eyes can grab on to parts of the image on the right, the one on the left makes my eyes tired and loose focus

>>106637858
>>106637941
hur dur im retarded, the left looks closer to what its meant to be and the right looks overly smooth and makes my eyes tired and loose focus

He's the guy who worked on the subsurface scattering used in the Call of Duty games. Those never get enough credit for what they pulled off on 8th gen hardware.

>>106637957
There's texture there that's not just compression that's clearly lost. Also a lot of the smaller brush strokes are just missing on the right.

Compression tech is so god damn boring and worthless at this point I have no idea why anyone thinks it's interesting. We're at the point where whatever is compressing shit is either just hallucinating details or omitting them completely.

>>106637957
It's a digital drawing. That's literally what it's supposed to look like.

>>106637786
>higher PSNR
>less high frequency details
what gives? could this be, uh how to put it?, they optimized for PSNR but in doing so ran into a limitation in using PSNR for quality comparisons?

>>106637786
>>106638068
>Source: [Zhang and Li et al. 2025]
I have worked with Chinese CS papers in the past and their metrics are inevitably faked or broken and their results not reproduceable.

>>106638129
Seems pretty diverse to me

>>106638068
it's clearly not though. You can see brush strokes missing in the reconstructed image compared to the jpeg. It's just missing detail and obviously rounded and smoothed.

I would point out the obvious but it's obvious so I don't have to.

>>106638262
Sure, but the jpeg isn't perfect either.
The reference image is in the top left the new method is definitely the better of the bunch.

>Two Minute Papers
>Videos are far more than two minutes long
Maybe this isn't the person to trust when it comes to data compression.

>>106637786
>JPEG
So how does it compare to codecs that aren't >30 years old?

>>106638262
It's not lossless compression you mong.

>>106637777 (OP)
checked
>>106637786
they are trying to make compressed pics look like ai slop so consumers can't tell and the push back against ai fades away

Seems weird to compare image quality for various formats by looking at an image in a single format. It's like being amazed by the image quality of an HD tv you saw in ad on a standard def tv.

>>106637786
preserving detail is a bitch; jaggies are preferable to blur

>>106638293
If feel like the jpeg is a lot worse than what it should be. They always do this shit. I seriously have never seen a jpeg that bad before and I'm a professional artist.

>>106638439
It's not just lossless, it has weird artifacts that just aren't blocky. As in, things are just being smoothed over and it's just losing detail to the point where it's just not any better.

>>106638509
Because the jpeg is 159KB, that part makes sense
They are comparing everything at ~160K

>>106638393
/thread

>>106638525
160KB isn't exactly a small file and that much compression for a 160KB jpg is... it's a bit much. I get they are trying to show off but that's not realistic at all.

png or death

>>106637786
>filter away al the detail
>still the same file size as JPEG

lol lmao

>>106638509
JPEG is Good Enough (tm) for personal use, the only people who care about muh ultra-efficient detal swallowing smearing codec are CDNs who pay top dollar for traffic

compression is for fucking niggers and if i compress something i do not give a shit about it in a first place, chasing compression standard or so called improvement is like chasing newest garbage bin holy shit i do not give a fuck kill Yourself

>>106638900
>top dollar
more like bottom pennies

>>106637786
OK, so it's better than JPEG (not even the best image codec available) for one particular image at potato compression ratios. Now compare it on an entire image set. I am sure it would probably sit the bed when trying to encode high quality images, and may even loose to webp and avif at high compression ratios. This is not a good paper.

>>106638856
based

>>106637777 (OP)
it was solved 30 years ago

what a time to be alive

>>106637786
kek

Gaussian splatting for 2D images has been around for well over a decade, this shader is originally from 2010:
>https://www.shadertoy.com/view/MdfGDH
and a more decent version from 2023:
>https://www.shadertoy.com/view/dtSfDD
Not a novel technique like the youtuber is implying and would be nothing special as an image codec

>>106637786
It is just easier to add a Neural Network in post to reduce noise.

>>106638068
JPEG artifacts are SOUL

>>106637786
That little triangle looking thing a little bit above the midpoint of the black square is a different color in both images

>>106637786
>smeared shit