Thread 105809679

Is VMAF a good quality metric for cartoon images? I want to up my image codec autism game. If you launch this windows script from the CLI directly it will encode a webp to a desired quality setting and spit out a vmaf score.

@echo off
set /p quality=Enter quality:
cwebp -m 6 -q %quality% in.png -o %quality%.webp
ffmpeg -hide_banner -i in.png -i %quality%.webp -filter_complex "[0:v][1:v]libvmaf" -f null -

I'm gonna try it out across different webp quality settings and try to match it with libjpeg-turbo.

>>105809679 (OP)
vietnamese male asian female?

I'll do increments of 20 until I reach 100. 0 will be skipped since I doubt anyone would use this irl.

Q20 Webp
VMAF score: 89.283047
24.6 KB file size
https://files.catbox.moe/njvjla.webp

libjpeg-turbo failed to reach this file size so no comparison to make.

>>105809687
vaginal male anal female

>>105809679 (OP)
Now do this with some more risqué pictures of Nobara

Q40 Webp
VMAF score: 92.205984
32.3 KB file size
https://files.catbox.moe/7p8phl.webp

>>105809950
libjpeg-turbo ALSO failed to reach this file size but fuck it, good enough for a comparison and VMAF correctly predicted that this looks like absolute utter fucking dogshit.

VMAF score: 43.681437

>>105809679 (OP)
complete fucking retard here: is vmaf really usable for static images? like isn't it tuned for video?
but besides, stop wasting your time on lossy webp, it makes no fucking sense to use, it's barely any better than mozjpg, either use that or go for jxl/avif, it's not like webp has great compatibility anyways and it's probably going to end up being a forgotten format nobody uses in X amount of years so not good to archive stuff either

>>105809992
why libjpeg turbo instead of mozjpeg/jpegli???
are you running a 386?

>>105809950
My queen...........

Anyways, I'll be using your research to know what settings to use in a gui

Q60 Webp
VMAF score: 93.893534
40.2 KB file size
https://files.catbox.moe/b16b7i.webp

>>105810003
>>105810017
Unfortunately 4chan will mess with JPGs encoded with mozjpeg/jpegli so it's not possible to easily use them for a comparison here. Anyway for the purposes of gauging whether VMAF is of any use for image quality comparisons, specifically cartoon images, I HAVE to use libjpeg-turbo because mozjpeg/jpegli might have been tuned to cheat in VMAF. Similarly VMAF cannot be used with JXL/AVIF since they're guaranteed to cheat in them.

I'm doing all this shit because BOTH SSIM and PSNR failed to predict quality of cartoon images with BOTH Webp and JPG.

>>105810147
I guess there's no point in making a libjpeg-turbo comparison to this since >>105809992 is pretty close to its file size...

Man, what is it with cartoons and libjpeg-turbo? It's getting fucking mogged here so badly I feel pity for it.

Q80 Webp
VMAF score: 95.547430
55.8 KB file size
https://files.catbox.moe/gp2gum.webp

>>105810224
Okay finally, somewhat of an improvement. Image quality is still pretty bad but it does seem like VMAF scores of above 90 and below 95 are "good" quality. Similar to what you would expect with SSIM so this is VERY promising.

VMAF score: 82.274176

WEW last quality setting. After this it's gonna be a guessing game trying to get JPG file size up high enough to match Webp's VMAF scores.

Q100 Webp
VMAF score: 97.304532
175 KB file size
https://files.catbox.moe/8ck7xp.webp

>>105810147
> I HAVE to use libjpeg-turbo because mozjpeg/jpegli might have been tuned to cheat in VMAF. Similarly VMAF cannot be used with JXL/AVIF since they're guaranteed to cheat in them.
oh so you're just that schizo again, ok...

>>105810351
Interesting. libjpeg-turbo was finally able to match Webp filesize and quality but only after an exponential increase in filesize.

VMAF score: 97.000437

>>105810278
>looks like completee garbage but mozjpeg is bad because... it cheats the scores ok, the point of lossy compression is to make the computer give happy number instead of making human happy!!!
no one gives a shit about your mental illness, this thread is completely fucking pointless, kys.

>>105810448
4chan messes with mozjpeg/jpegli uploads you fucking dumbass.

>>105809679 (OP)
even netflix, the jews that literally created vmaf refuse to use it for still images and stick to ssim instead: https://netflixtechblog.com/avif-for-next-generation-image-coding-b1d75675fe4
meds, please.

>>105810477
it losslessly reencodes them and that causes those very optimized jpegs to become bigger yes...
so your solution to this is to upload crappy output from a shitty encoder that hasn't been relevant in 20 years... how the fuck does that make any sense?
1. it's a "problem" that only affects 4chan
2. you can just use a lower quality setting for mozjpeg to aim for the same filesize
I have no idea what your end goal is here.

>>105810479
What do you think VIF is genius? BOTH SSIM AND PSNR COMPLETELY AND UTTERLY FUCKING FAILED TO PREDICT CARTOON IMAGE QUALITY NOT JUST IN WEBP BUT IN JPG AS WELL.

fuck

https://github.com/Netflix/vmaf/blob/master/libvmaf/src/feature/vif.c

>>105810495
I don't trust what 4chan is doing. Maybe I'll redo this whole thing and upload the jpegli versions in a zipped up folder in the future.

>>105810495
>it losslessly reencodes them and that causes those very optimized jpegs to become bigger yes...
and besides... it doesn't
the filesize displayed is just wrong, my guess is that the code throws away the optimized jpeg and sticks with the original if the transcode makes it bigger (but the filesize display stays the same as the optimized jpeg that was discarded)
if I save my 72 kb jpeg, it's still 55 kb like it originally was
so yeah you are doing all this schizo shit for what's clearly just a bug that any person with a working brain will figure out in 30 seconds
I probably reached the same conclusion in the past but had just forgotten about it, and you ignored that fact like every single response you get
I have no idea why you even do these threads, they focus on pointless shit that nobody cared about, and when people try to make you reason and explain why you're fucking retarded you just never acknowledge anything, it's tiring and I'm just a retard for even trying to """help""".

>>105810537
You still trust 4chan that much after that sharty hack? LOL

>>105810532
>What do you think VIF is genius?
...a version of vmaf that's optimized for still images and that you aren't using? it literally has a different name, i wonder why...

>>105810555
>You still trust 4chan that much
...reading the file size of the local file on my local machine is somehow "trusting 4chan"?
...the sha256 is literally the fucking same as the file I uploaded

>>105810583
Not good enough for me but you do you boo.

>>105810602
>the sha256 being the same is not enough evidence that the file is the same
ok, i'm done, won't interact with your ragebait anymore, bye.

>>105809883
JEEEESUS christ so after 10+ attempts I finally got a similar VMAF score to a Q20 Webp.

24.6 KB webp divided by 62 KB libjpeg-turbo = 0.4 so Webp achieves 60% better compression efficiency than libjpeg-turbo.

I'm not sure if this sub-90 VMAF scores works similar to SSIM to denote low quality but it does seem like it.

So far all this data amounts to this basic graph. I guess the next step would be to painfully match these VMAF scores with libjpeg-turbo and make another graph showing the file size increase compared to Webp. So far I got Q 20...

Maybe VMAF scores could become the Y axis labels? Something like this I think.

Getting the other libjpeg-turbo results is going to take forever thoughbeit. Can any of you fucking nerds help me speed this up somehow? I only have a high school level windows batch scripting proficiency.

I'll do jpegli as well even though I suspect it may cheat in VMAF. I mean google did release this thing in 2024 and VMAF was released in 2016 so there was plenty of time to sabotage VMAF scores but whatever.

https://opensource.googleblog.com/2024/04/introducing-jpegli-new-jpeg-coding-library.html

https://netflixtechblog.com/toward-a-practical-perceptual-video-quality-metric-653f208b9652

>>105810003
>it's not like webp has great compatibility anyways and it's probably going to end up being a forgotten format nobody uses in X amount of years so not good to archive stuff either
Both modern Android/Windows/Mac natively support webp and older versions can be augmented to support it via third party image viewers/editors. Pretty much all web browsers support it too.

I really doubt it's going away soon because even though AVIF and JXL both outperform it (at least in lossy compression), Webp still has the advantage of fast software decode speeds which matters for all the low end electronics out there.

JPG might have become better but it still doesn't support transparency or animations so it's still practically an obsolete soviet-era technology.

>>105809679 (OP)
VMAF is for video you idiot.
>>105810003
Still less retarded than OP, because you're correct.

>>105811809
I'm trying to find something that will work with cartoon images. VMAF seems to be working but if there's anything better than jerry rigging a video quality metric to detect image quality, I'm all ears man.

https://desuarchive.org/g/thread/105805949

>>105811907
Real? Subjective double blind ABX testing with a sufficiently large group of people. That's the real answer.
You could try Butteraugli. Guess what? jpegli and cjxl were tuned based on that metric. You're fighting a losing battle idiot, a these different encoders were TUNED BY DEFAULT TO GAME SPECIFIC METRICS. IIRC for cwebp they tuned for PSNR or SSIM.

>>105811934
I'm not testing modern formats, just plain old Webp and the more ancient JPG image format.

I got a score of 6.1268320084 for the 62KB JPG in >>105810672
vs a score of 8.4359798431 for the 24.6 KB Webp in >>105809883

I would definitely not rate the JPG as better than the Webp so I can't-believe-it's-not-butter is an utter failure like SSIM and PSNR for cartoon images as well. I'm not giving up yet thoughbeit.

>>105812049
Those are both awful scores, and both images look like shit. It's working just fine. It's also meant for high quality comparisons.

>>105810532
very cute.

>>105812109
This is webp so "high quality" isn't even a consideration here lol. Just good old "good" quality you get with a quarter pounder from the clown vs a cheeseburger from a 5 star restaurant.

Fuck, I'm too deep in the shit to quit now. Just one more set of results.

anon's new fixation... very cute enthusiasm.

WELP there you have it folks. At least when cartoons are involved Webp achieves around 50% better compression efficiency against the old libjpeg-turbo JPG encoder and around 30% better compression efficiency against the new jpegli JPG encoder. So while jpegli has made very impressive gains, I don't see Webp going away in the near future.

>>105813224
Kinda strange seeing Webp compete this well in anime content. I always thought this was AVIF's main thing.