This is How Google’s Phone Enhances Your Photos

📝 The paper "Handheld Multi-frame Super-resolution" is available here:
sites.google.c...
❤️ Pick up cool perks on our Patreon page: / twominutepapers
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
313V, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Anthony Vdovitchenko, Brian Gilman, Bruno Brito, Bryan Learn, Christian Ahlin, Christoph Jadanowski, Claudio Fernandes, Dennis Abts, Eric Haddad, Eric Martel, Evan Breznyik, Geronimo Moralez, Ivelin Ivanov, James Watt, Javier Bustamante, John De Witt, Kaiesh Vohra, Kasia Hayden, Kjartan Olason, Levente Szabo, Lorin Atzberger, Marcin Dukaczewski, Marten Rauschenberg, Maurits van Mastrigt, Michael Albrecht, Michael Jensen, Nader Shakerin, Owen Campbell-Moore, Owen Skarpness, Raul Araújo da Silva, Richard Reis, Rob Rowe, Robin Graham, Ryan Monsurate, Shawn Azman, Steef, Steve Messina, Sunil Kim, Taras Bobrovytsky, Thomas Krcmar, Torsten Reil, Zach Boldyga.
/ twominutepapers
Splash screen/thumbnail design: Felícia Fehér - felicia.hu
Károly Zsolnai-Fehér's links:
Facebook: / twominutepapers
Twitter: / karoly_zsolnai
Web: cg.tuwien.ac.a...
#NightSight #GooglePixel

[@Favmir]

"This doesn't use any neural network technology" *Top 10 Anime Plot Twists of All Time*

[@martiddy]

I know right?, I was like: "I wonder what kind of AI structure they used", and when Károly said that they didn't use any neural network whatsoever and is hundreds of times better than other previous methods, that totally blew my mind

[@kevincui1631]

If you just know a little bit about machine learning, you know neural network and machine learning don’t solve this kind of problems.

[@Guztav1337]

@Kevin Cui , Except they did, look at previous videos. They got really good results with machine learning.

[@AlxM96]

@@kevincui1631 look up ESRGAN

[@ChristopherGray00]

hahahaha funny!!! definitely not an over saturated dragged out meme literally said over a million times already!!! your so funny!! im a normie simpleton who enjoys dead memes!! laughing out loud crying emoji x10

[@itachi2011100]

Two minute Papers: "Hold your papers!" Me: *holding tightly* Two minute Papers: No Neural networks *Papers blowing away*

[@LaLogic2]

You want me to help you pick those up

[@drastelne]

"You can't just zoom in and enhance, adding more detail from nothing" Google: Hold my algorithms

[@michac3796]

Start enhancing 80's TV shows. Thank you.

[@THarshavardhanReddy]

it's not from nothing though, is it? they are going to use multiple frames.

[@dsp4392]

BUT WHAT IF IT DID USE NEURAL NETWORKS?! *Wild investors appear all over the place*

[@Keldor314]

I can think of a use for neural networks here, actually. Temporal reprojection. Imagine a scene where you have a number of moving objects like cars, which don't stay stationary with respect to the background. If we can understand exactly where they've moved in each frame, we can combine the data from each image anyway while still being reasonably accurate. But while cars might be easy for a conventional algorithm to understand the motion of, there are many far more difficult objects in real life, such as perhaps tree branches. Matching arbitrary overlapping patterns while keeping track of occlusion sounds like a *very* hard problem for anything but a neural net.

[@MobyMotion]

Love the elegance of taking a weakness (shaky hands) and taking advantage of it. Wait?? A Google image processing paper that doesn't use machine learning? Gonna have to rethink my obsession with neural nets. Fascinating to see cases where hand crafted algorithms still outperform ML. PS: Please can we see more of Lisa in future episodes

[@TwoMinutePapers]

She is planned to make another appearance in a few episodes. Stay tuned!

[@diduseethat3191]

Yes but now imagine using this method combined with neural nets. Also who knows if this isn’t an already AI developed technique;)

[@kevincui1631]

You know nothing about ML if you think ML outperform any “handcraft” algorithm. There are plenty of algorithms just can’t be done by ML.

[@MsZiomallo]

@@kevincui1631 but ML is nearly impossible to beat if task involves imagination. The best denoising algorithm is... yes, imagination. The best super resolution algorithm is... again in our brains. Machine learning just tries to learn from The Master (imagination) and mimics its behavior. So I'm sure that this algorithm is perfect starting point - I think that if we involve machine learning, results would be even better than that. Same with denoising - ML-based algorithms do that task better than classic algorithms because they know how to denoise particular parts of the image because they've seen similar parts in learning dataset. Isn't that human-like? We also are seeing the world around us and based on that, we're upscaling and denoising images in our brains like no computer algorithm ever could, because we have real world images remembered in our brains as a reference.

[@diduseethat3191]

Ziemowit Zabawa - Important to note that so many are overlooking the significant importance of todays image recognition is truly magnificent already enough to start loading historical document/visual into the system to teach what people like/love visually. Human creativity as we know it today eventually will be taken over by AI therefore the truly useful advice a wise man once said - lear to code

[@raphirau]

Awesome to see that there are still solutions being developed, not jumping on the neural networks hype-train! Additionally beating other learning use-cases is pretty amazing!

[@WangleLine]

I'm surprised it yields clearer results AND is faster than previous neural network solutions. Amazing :D

[@13x666]

“Hype train” is not exactly a great way to describe NNs though. NNs can do some extraordinary stuff, and the hype is just a result of that. It's not just a buzz-word, y'know.

[@CuulX]

It's cool that it works now. But NNs will beat it soon. Progress in NN and hardware is more beneficial in the long run. That's the bitter lesson. But it's actually pretty sweet when you embrace it because of how it helps in every category we can define 'practice' or training data for.

[@diduseethat3191]

Hello, it is Google:) One of the largest researchers in the fields of AI, I’m sure there’s more behind than meets the eyes. ML/DL etc supposed to develop tools like this I’m just not sure where we’re at with its development.

[@lightsounddevice8983]

I hear you on the "hype train". Some new graphics coders out there don't even know how to do a simple FOR loop on image pixels. Find #FF0000 in pixels? Needs AI.

[@CHL41993]

This is real signal processing, not some random AI blackbox.

[@yigitpolat]

when are you guys going to disappear?

[@blueberry1c2]

I remember a video from Lindybeige postulating that super-resolution images can be made using a similar method. He made the analogy of a screen with lots of tiny holes through it: if you stand still you can hardly see anything behind, but if you slowly walk along the environment behind the screen becomes much clearer.

[@Usernamewhatever1443]

my first thought too

[@blueberry1c2]

@@zaandam0172 if you have a microwave with that sort of pattern on the glass you can try moving your head around vs keeping still

[@blueberry1c2]

The holes represented pixels on a ccd i guess. He isnt really into cameras and technology he was just disproving csi movies' magical enhancing technology

[@s.m8766]

Isn't this the process used for creating astrophotos long since? Stacking of multiple frames and also drizzling to increase resolution. Have a look at the software "Autostakkert!3" which is used to make planetary images from video frames.

[@gasdive]

Looks the same idea to me. Slightly different details. I've been using it for years. See also deep sky stacker and photoacute.

[@s87343jim]

We use the same idea for landscape, architecture works too

[@AdiCherryson]

As I've understood the idea isn't new just the algorithm

[@ShahriarShanto]

I suggest you guys read this: ai.googleblog.com/2018/10/see-better-and-further-with-super-res.html

[@TheHuesSciTech]

IIUC, astronomical scenes are much easier because the scene is just a bunch of point sources (stars). Whatever the simplest, pointiest explanation for a sampling of collected, blurry frames is, is probably the right enhanced image. Except for nebulae and such, which are fuzzy, and therefore don't need to be enhanced to look right. Also, the "movement" in astronomical imaging is much simpler (no parallax changes like with a human hand holding a phone with nearby objects.) So all in all, although there might be superficial similarities, the results of replacing the phone's Night Sight software with Autostakkert3 would be absolutely laughable. Also, the video shows plenty of comparisons with competing algorithms, so obviously there is prior art. The point is, this is considerably better.

[@feature-creature]

Can't wait for this sort of thing to be eventually applied to video restoration technology for, say, things shot on or only available on VHS.

[@bryanotero123]

Brilliant

[@russianbotfarm3036]

In the next: "They Shall Not Grow Old", maybe: kzbin.info/www/bejne/f6PEk36BbqebmtU

[@dsp4392]

If anything, persistence of vision pretty much already does this for video, so there would be little to gain by artificially "upscaling" video frames. You're already seeing all the info there is. That said, neural networks can already be used to "dope" images by introducing detail that should statistically be there. For instance, take heaps of high res video, make a low res copy, train a NN using the low res data as input and high res data as output, then use the model on old footage.

[@pandarzzz]

I learn greatly from Two Minute Papers... Subscribed! ✅

[@timbomb374]

Using multiple images taken with shaky hands to make one high quality one, that's some genius out of the box thinking.

[@Bolt6265]

I've been thinking about this exact idea for yeeears, It always seemed obvious to me that taking a bunch of pictures from slightly different positions/angles gives you added information that could be combined to reconstruct a more detailed image. Nice to see it finally being brought to full realization. Surely this technique could also work on any existing video footage to create high resolution stills from existing video?

[@Guztav1337]

We are so many on the earth, that we are never alone with our thoughts.

[@russianbotfarm3036]

I wonder whether this can reconstruct a digitized-out version. You know, like where they have someone talking but his / her face is pixelated? I've often thought that there may be enough information in the different, changing 'big pixels' to reconstruct the original face.

[@Bolt6265]

@Jeremiah Bullfrog Ohh that's an interesting thought. Quite interesting, and spooky. I'd think you'd probably need to modify or tune it specifically for doing that but might work to some degree.

[@gasdive]

It's been around for years, mostly in the astro photography world. I've been using it for about a decade. This seems to be a refinement. It's already commonly used to turn video into high resolution stills.

[@Guztav1337]

This is indeed a refinement, a very good one at that.

[@kallamamran]

You know you have a great idea when you have an idea of how something could be done and then hear about someone who have actually done it! :D

[@rachelslur8729]

The camera could slide-off it's infra-red filter to get even lower noise low-light images with all the extra infrared photons.

[@Max24871]

You would also get a lot less atmospheric haze by doing that!

[@rachelslur8729]

@@Max24871 👍 Yes and a post processing color curve would remove weird ir redshift, while reducing image noice. It's nice to see a fellow smart person.👋

[@kalazakan]

​@@rachelslur8729 Jesus, how pretentious...

[@rachelslur8729]

@@kalazakan 🤣

[@AXYZE]

@@rachelslur8729 And how would you differ bricks and autumn leaves via "post processing color curve"? Then you have animal eyes, human skin etc...

[@יובלהרמן-ח2ד]

I was using Hugin to align and make super res photos for years but they definitely did something better cause I take more like a couple of minutes

[@YGODueltainer]

Google's night sight is like pure black magic, turning low light photo to almost day light shot.

[@erikziak1249]

This method might be also very useful to boost dynamic range. Nice.

[@MsZiomallo]

Well, they're already doing that since Pixel 2 :) Haven't you heard about HDR+? It also has its own papers. Just search a little bit on the internet and you'll find them ;)

[@erikziak1249]

@@MsZiomallo Thanks. I am too old for that. Just last week I took my old 35mm Praktica VLC 3 with good old Carl Zeiss Jena lenses, put a Fomapan 100 film in, shot some pictures, developed the film in Fomadon R09 and scanned the negative. The dynamic range was well above what I needed. I just wish I had a better scanner for 35mm film.

[@c64cosmin]

This is alike how telescopes use to gather more information to resolve clearer images, not really novel IMHO. But great that it can be used on a handheld device with movement compensation.

[@peetiegonzalez1845]

I did exactly this algorithm 25 years ago. Not using neural networks but using exactly the same principle by averaging out visual data based on relative location within the assumed fixed scene, over several moving frames. It was just a novel exercise back then. I'm surprised this paper is considered ground-breaking.

[@peetiegonzalez1845]

I mean... not to belittle the authors. There's a difference between mental masturbation and producing a publishable work. Well done to the authors and I hope more people push their little exercises to a publishable result that will inspire others.

[@alanjenkins1508]

Now you just need a version that improves moving objects within a static camera viewpoint. Useful for improving surveillance camera images like in the movies.

[@thecapacitor1395]

Riminds me of Registax.

It’s not entirely new. We used this technique of stacking frames 20 years ago in astrophotography, especially for making pictures of planets like Jupiter and Saturn, using a webcam with the lens removed instead of the eyepiece. I don’t remember the name of the software we used for stacking, aligning and averaging the frames, but it worked great. It’s indeed magic to see such a high quality image arising from a bunch of noisy frames :)

[@OutOfNamesToChoose]

Lisa, your chihuahua, is adorable

[@hvgpilaatkaiken2300]

Question: Combining multiple discrete images or scanning over continuous open-shutter stream from the sensor? Assuming one can pull an open-shutter stream from the sensor, wouldn't that data be of a higher quality in contrast to discrete signal provided by the same feed preprocessed to arbitrary chunks the size of the shutter speed? Or would that be inconsequential enough to not be worth losing hardware acceleration over (if it's even feasible processing-wise)?

[@U-D13]

The shots of architecture at about 2:50 (and another shot a bit later) are from my hometown. Cool.

[@Quoutub]

The theory(!) makes sense, the problem is that Google's zoom doesn't work. Theory and practice are different things. It combines more frames when you zoom, this leads to less noise, so this leads to more detail, if(!) the light conditions are not perfect. But if the light conditions are perfect and there are no shadows, then it's hard to see any difference at all (when you pixel peep and compare the result with a single frame dng file or when you compare the zoom result with an older Google phone like the Nexus 5x, one doesn't really see a difference). I have a Pixel 3a and tested the zoom extensively over multiple years (I already had compared results when the Pixel 3 was released). In perfect light conditions it's really not better than traditional digital zoom. Also note that Google does combine this technique with machine learning zoom, which leads to smoother contours, but doesn't add true details as machine learning only guesses. It can actually even cause very wrong details and perform worse than traditional digital zoom. I was very disappointed that this was just a marketing stunt, but that's the truth. When they released the Pixel 4, they continued with their marketing lies and indicated that the zoom is comparable with 3x lossless zoom and this was again a lie. I even calculated that this is even physically not possible due to diffraction as the lens of the Pixel 4 "tele" has a too small entrance pupil. (No algorithm can prevent diffraction, this would violate quantum physics).

[@henrykzajac8001]

After all algorithm didn't enchanced photo. It requires many similar frames to do so. So the begining is a bit confusing. You can't boost quality of photo already taken. You can take better photo with algorithm built in camera. (or eventually you can use it for video)

[@p0rt3r]

I think the new thing here is where and how they apply this technique, right? The technique itself sounds like image stacking.

[@RAJATTHEPAGAL]

Google have done something like this before also. And especially not with neural networks. Your channel covered it. 😁 .... Awesome stuff .... Will check out.

[@artmusic6937]

I want to say that this technique with super-resolution was already studied by Patrick Vandewalle et. al in 2003, 2005 and 2007. He already mentioned the use of this in digital cameras. Pls add the name of this man he deserves this. Check the papers link.springer.com/content/pdf/10.1155/ASP/2006/71459.pdf

[@dislike__button]

Cool stuff. Can I let go of my papers now?

[@TwoMinutePapers]

Hold carefully until the end of the video. After that, yes, you can let go. Gently. :)

[@castoru3398]

N̴o̷.̶ ̴Y̵o̴u̴ ̸m̴a̷y̴ ̵n̵e̵v̶e̵r̷ ̴l̵e̶t̴ ̶g̷o̵ ̸o̷f̴ ̵y̸o̴u̴r̶ ̴p̷a̸p̶e̵r̷s̷

[@maxfmfdm]

Nice. This has been done in astronomy for some time. No reason why it cant be done by our phones.

[@LegoEddy]

Next phone is gonna be a pixel. It's just so exciting to see how they put the latest research results into their phones, even after release as software updates!

[@kizonyoutube]

I've thought a couple times about using more photos to use to make a clearer one, but for obvious reasons I didn't have a clue on how to execute it.

[@MrEazyE357]

Whenever I take someone's picture for them, I always take at least three. That way they can choose the best one.

[@PennyHerbst]

Sometimes when I'm out in the city, I'm thinking why is the starbucks always that crowded

[@noutram1000]

Isn't it just collecting more light (more frames), you get the same result with a still camera -if you leave the shutter open longer. OK, there's some smarts in there to adapt for movement of different types but the more photons you can collect the higher the information to noise ratio will be...

[@michelles.3835]

This was a genuinely interesting video.

[@secnep]

I have a Google pixel 2 xl myself, this is so interesting I love it.

[@sefrautiq]

I wish if Google camera's Nightmode & HDR had the same posibilities in quality and editing as in the RAW format. When I made my first shot in raw in lightroom mobile and then found out how I can edit it, I just.. I just barely touched any stock or camera app ever since.

[@markbakker5214]

Does this mean that making more and more pictures enhances the resolution indefinitely or is there another limiting factor?

[@hyunsunggo855]

Not using neural networks = Just waiting to be applied to a neural network to be much faster and effective.

[@Hacks4AllVideos]

This is awesome.. thank you for sharing this.. Congratulations!

[@Astronomy_Live]

Sounds like a modern version of the drizzle method developed for enhancing the resolution of Hubble Space Telescope images.

[@99bits46]

Apple in 2022: We're introducing Advanced Liquid Sensor technology that will enable super-sharp Images. Something that's never been done before. Apple Consumers: Hold my 1800$

[@zenonuchida263]

What a time to be alive!

[@DigitalicaEG]

ENHANCE!!

[@dawg_gee_man]

You need to find out what algorithm the Huawei P30 Pro uses and compare that!

[@calorus]

Awesome review.

[@sanjaynagarimadugu6108]

What a time to be alive

[@spaceowl5957]

Doesn't temporal anti aliasing do a very similar thing in real time?

[@seadve]

Computational photography is the future.

[@jaredtweed7826]

Nice! This is the first video I saw by you. I subscribed.

[@hippopotamus86]

I had this idea years ago, just had no idea how to do it.

[@dsp4392]

So you didn't have the idea?

[@hippopotamus86]

@@dsp4392Yes I did, I thought up the idea of having several photos of the same picture and using the information from each to create a more detailed image.

[@epsospremium6088]

*Impressive results.* Good for phone photography. This will be also very useful in Ai object recognition for darker images. Many people will benefit from this, because it does not require to buy a new phone, just an app upgrade.

[@wanhl2440]

neural network is useful when you have only one image input. Stacking the photos give them effectively more exposure, i.e. more total light gathered by same sensor size. It is not fantasy but the law of physics. The camera sensor of google pixel 3 is 1/2.3 inch, close to early compact camera. However, if fact there are no compact camera with this sensor size matched pixel 3 night sight in low light. They did not have the processing power to stack photos like that

[@prabhakarmishra2182]

Great work both by Google and you

[@SocratesAlexander]

That's what our eyes and brain do all the time. You know about the micro-movements of the eye to boost the resolution.

[@lolerie]

Yeah)) en.wikipedia.org/wiki/Microsaccade

[@leonc4653]

yeah for image stacking,

[@cthulpiss]

So.... is it hardware dependent? Or can you make it into app for any decent phone?

[@McGeri]

Köszönjük, nagyon eposzi.

I remember when nightsight was just suddenly there...

[@crackedemerald4930]

Could this technique be used for astronomy?

[@serpinio]

Lol. Median averaging of a series of shots with automatic discarding of statistically irrelevant areas is what they call breakthrough nowadays? Breaking news! Neat Video has been doing this for AGES with its "temporal" method, allowing to achieve the same results thanks to simple median averaging techniques.Welcome to the woooooorld of tomorrroooow! Still though, I have to admit it's cool to see they honestly acknowledge this is a fast and simple method and not some NEURAL NETWORKS MAGIC. I respect that.

[@PresidentialWinner]

Subbed after your Open AI video. Now i'm glad i did.

[@musliksolihin8960]

So it's basically temporal anti aliasing?

[@MsSebaastian]

WE NEED MORE VIDEOS OF THAT CHIHUAHUA !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

[@phreakinpher]

Can anyone explain what's going on in the background of the live photo? Its like dancing around strangely.

[@zac9311]

This one kinda blew me away

[@seventysevencats]

The software gets better and better but the hardware is still a limitation. As they make the phones slimmer, they get less and less room to fit lenses - even if they can now use folded optics.

[@invertedgames7993]

looks like Cinema Sins is gonna have to remove some sins.

[@m_sedziwoj]

So put it to DNN and get even better results ;)

[@RAJATTHEPAGAL]

I have worked with Super resolution networks before 😅 this work is far more impressive given it's speed + reliability. DNN can , will fail more importantly u can make it fail. Super res network in one shot are doing great no doubt but the cost at which they do it makes them not so portable. So i think more work like this should be encouraged. Less neural network more understanding the problem. 😀

[@3liMoe]

i got Samsung Ad after this video

[@AsmageddonPrince]

Well... now I just feel silly thinking I was the first to have this idea, jeeze

[@lolerie]

Actually i thought about it the first time i knew that our hands are shaking)) And i knew about 3d glasses before, so... I just put 2 and 2.

[@kashparty]

I got a Pixel ad for Night Sight

[@abdohossan]

I don't even watch movies but I can imagine some CIA lady saying "ENHANCE THIS IMAGE JOHN" or smth.

[@truelies5431]

I'm lovin' it!

[@Jacob-yg7lz]

Lindybeige was right!

[@geevarghesegeorge1424]

So basically a HDR sequence processed differently.

[@alexline4131]

I like Lisa

[@baseladnan77]

Brilliant! Is the tool available to the public? I would like to improve some of my photos x).

[@MatesMonchis]

3:08 DeepJoint LMAO

[@qweqwe_Flinstone]

EMERGENCY! EMERGENCY? hmmm seems working fine (first photo on the begining)

[@norsemodder]

It probably does say use a '?' based on the fact that the title says 'Employee first aid station' and that there is another line of text below. I would assume it said something like: Employee First Aid Station EMERGENCY? VISIT THIS ROOM That's just how it appears to me, though.

[@dragonskunkstudio7582]

Isn't this temporal denoising?

[@WildAnimalChannel]

This isn't new. I used to combine frames from my webcam to get sharp images like 20 years ago! I mean, the sub-pixel stuff and removing the shake is probably not that complicated either.

[@Guztav1337]

Well, no other smartphone has had this before. And this quality is unseen from other manufacturers. Obviously, it is new.

[@pixelfodder]

E N H A N C E

[@toyuyn]

This seems kinda similar to beamforming.

[@pencrows]

Tripod intensifies

[@SoftYoda]

Will there be one day all technology of smartphone camera to real DSLR ?

[@dsp4392]

I don't know, but we have the technology to form complete sentences RIGHT NOW!

[@SoftYoda]

@@dsp4392 sorry

[@alex_on_the_web]

If the paper is out, and the method sounds so straightforward, is it a right guess, that Night Shift make it to the revolutionary 2020 iPhone?

[@ravenmooore]

Z O O M. E N H A N C E.

[@doku7335]

I thought "improving resolution by aggregating information from several images" was original idea of Super Resolution, and the fact that the newer methods were called "Single Image Super Resolution" was showing how much the algorithms have improved - we don't even need multiple images anymore. But now we go back full circle and return to the case of several frames, taking this to the next level again.

[@Phobos11]

Eldar not really, stacking is a very old idea, nothing new here

[@johneygd]

Such technology already exists for a long timme, sony uses this method to filter out blurring,smeary rolling shutter arifacts ,noise, ,shaking, so again this is nothing revolutionary at all,

[@nikch1]

This comment section basically: "I can't believe/I'm surprised it took them this long to do that because in photography we've been doing this forever." etc If you all find this so simple, easy, and not even worth your time why haven't you done it or contributed to such project before then? I feel like people who follow this channel are more and more the type of arrogant "I am very smart and I must show it to everyone" person.

[@W00denCube]

EMERGENCY?

[@mathnerd314159265]

> Two Minute Papers > 4:37 Something is not right here...

[@1KeyJee]

AI≠better

[@15DJONYX]

Me: #teampixel

[@TechInspected]

Search for Waifu2X. Thank me later =)

[@AManKumar-pz6et]

When Google is going to upgrade their pixel phone with this feature? I am so excited.

[@ManomiiFox]

there's a similar software out there called Waifu2x that does a similar thing: kzbin.info/www/bejne/g4ezo3mLa8SGe6M but unlike this one, it's actually usable by anyone. ONLINE VERSIONS: (from best quality/ fewer limitation to basic) waifu2x.me/ waifu2x.booru.pics/ waifu2x.udp.jp/ PC Versions: github.com/lltcggie/waifu2x-caffe github.com/tanakamura/waifu2x-converter-cpp (works on AMD cpu)