https://www.dpreview.com/news/0501469519/google-researchers-detail-new-method-upscaling-low-resolution-images-with-impressive-results [p] * * Facebook Twitter YouTube Instagram Sign up for our weekly newsletter! Submit a News Tip! Reading mode: Light Dark Login | Register [ ] [] NewsReviewsArticlesBuying GuidesSample ImagesVideosCamerasLenses PhonesPrintersForumsGalleriesChallenges Best cameras and lenses [logo-v3] Now reading: Enhance! Google researchers detail new method for upscaling low-resolution images with impressive results 241 comments 241 Enhance! Google researchers detail new method for upscaling low-resolution images with impressive results Published Aug 30, 2021 | Gannon Burgett Share Tweet [Google_Researchers_SR3] Cascaded generation of unconditional 1024x1024 faces. Thanks to CSI, as well as plenty of other crime dramas, the phrase 'Enhance' has taken on a life of its own as a tongue-in-cheek way of attempting to digitally extract extra information from low-resolution images that simply isn't feasible in real-world situations. Or is it? A new blog post on the Google AI Blog showcases a new technology its developed to upscale low-resolution images with incredible results. The blog post, titled 'High Fidelity Image Generation Using Diffusion Models,' explains how Google researchers have developed a pair of AI technologies that can take a low-resolution image and steadily increase resolution through selective destruction and reconstruction of the original input image. [CleanShot_2021-08-30_at_10] The before and after of increasing the resolution of the portraits from 64 x 64 pixels to 1024 x 1024 pixels. The first component of the process is Super-Resolution via Repeated Refinements (SR3), 'a super-resolution diffusion model that takes as input a low-resolution image, and builds a corresponding high-resolution image from pure noise.' In essence, this model applies pure Gaussian noise to a low-resolution image before using noise-reduction technologies to effectively reconstruct a nearly noise-less image that's four times the resolution of the input. The researchers then use Cascaded Diffusion Models (CDM) to intelligently apply Gaussian noise and blur to the output image before repeating the process again. This technique, which Google calls 'conditioning augmentation,' improves the image quality to the point that it surpasses current AI upscaling methods, which include BigGAN-deep and VQ-VAE-2. According to Google, this new technology 'achieves strong benchmark results on the super-resolution task for face and natural images when scaling to resolutions 4x-8x that of the input low-resolution image.' As visible from the above illustration, this means a 64 x 64 pixel image can output an impressively clear 1024 x 1024 pixel image. [Google_Researchers_SR3_II] Super Resolution results: (Above) 64x64 > 512x512 face super-resolution, (Below) 64x64 -> 256x256 natural image super-resolution. Google researchers say the technology '[pushes] the performance of diffusion models to state-of-the-art on super-resolution and class-conditional ImageNet generation benchmarks' and notes they're 'excited to further test the limits of diffusion models for a wide variety of generative modeling problems.' You can read the entire article on the Google AI blog. Tags: ai, google, research, technology, upscaling View Comments (241) Comments All (241) Most popular (15) Editors' picks (0) DPR staff (0) Oldest first epambos An analogue photograph bears the unique characteristic of an index as visual sign. The coming of digital photography has lead to many discussions about a severance between image and world, but one could advocate that the spirit of image creation even during the digital era preserved this connection or representation and reality. In this particular case though we have an example of a re-creation of an image that prioritizes cognitive reception rather than the faithful reproduction of wordly things. It is a shift in perception of what an images does (not dissimilar to Huawei offering to add a detailed archived image of the moon in your nightly photographs of the sky)! Like 0 30 min ago*permalink BasilG Looks very interesting for photography (i.e. uprezzing for large prints), but it's obvious that the detail is made up - made up well, but still, made up (look at the Leopard). Like 0 50 min agopermalink Mutovkin I experimented a while back with Topaz Gigapixel AI to "upscale" scanned photos and results are definitely not impressive - barely producing any additional detail. So artificially generating low res images from high res digital originals is not the same as say upconverting old photographs. That's the area I would be excited to see having more research done. It would be great to have a good collection of algorithms to process old photos and enhance them automatically removing dust, scratches, increasing resolution and optionally applying user guided colorization. Like 0 2 hours agopermalink stevevelvia50 https://youtu.be/TyNdi8y619o Like 0 1 hour agopermalink Hasa In some of these tests, that I have previously seen others make, they construct the low-rez image by down-sampling a high rez image. It is most probably not the same as starting out with a true low-rez image based on an imager with few pixels. So I would be careful not to overinterpret the results. Like 1 4 hours agopermalink JerryWilliam The term "High Fidelity" is thrown around rather loosely here. as... "Abstract art can be judged by any number of criteria, but resemblance to the original scene is not one of them, any image can be processed to appeal to an individuals taste, indeed there is little point in going through the effort in processing an image you don't want to look at, but whether or not the image looks better or worse than the original it is not TRUE to the original and cannot be considered a high fidelity reproduction". Paraphrased for context from Stereophile Vol. 1 issue 4, March/April 1963, by J Gordon Holt. Like 0 5 hours agopermalink Kaqpa What about down res? Can this virtual tool make my 8x10'' drum scanned slides look like 0,11mp from a Kyocera VP-210? Like 0 6 hours ago*permalink OneLeggedCat So cool! Soon, we won't even need the camera. Google AI can just make it all up for us! BTW, I look like Brad Pitt, so make sure to get that right when you are fabricating my image, Google. Like 0 6 hours agopermalink Maxbogue Wondering what it produce if you digital zoom so much until you are left with only one pixel picture... Like 0 7 hours agopermalink Istina Out of curiosity, does anybody has experience with topaz? Is it any good? Like 0 7 hours agopermalink kodacolor Gigapixel AI is quite good at enlarging images, in stages, at a certain point it can't enlarge any more and the more enlargement done the smoother the surfaces are in the image. Sharpen AI is OK, but certain types of blur it cannot sharpen; usually camera shake, camera movement, or soft misfocus. If the images are just off a bit i can make them quite usable. The biggest issue even on a very fast multiprocessor workstation is that Topaz is slow, quite a bit of culling needs to be decided on before processing as averages a couple minutes per image or more. Like 1 6 hours agopermalink Hasa I can make quite nice uprez from x2 linear from 10 Mpix with a great lens to 40 Mpix. I have also "saved" portraits of family that were slightly out of focus. I routinely sharpen up eyes and lighten the iris - using "Portrait Pro". However Topaz has examples of up-rez x6 and that definitely does not work for my images. About 2 to 3 times is the limit and useful for saving a bird image that is heavily cropped. Like 0 4 hours agopermalink centerpunch Topaz Gigapixel is totally amazing, almost unbelievable, on some photos. You can download the full program and use it for 30 days free to see if it works for you. It does need a good video card. Like 0 2 hours agopermalink filipradich Gaming in 16k coming in the future hah, derived from a full hd stream. Like 0 7 hours agopermalink aidaho Now that does look like something mildly interesting as opposed to Topaz AI. Would be fun to watch how this perform in a new wave of remastered old footage. Like 0 8 hours agopermalink SidePod why does that leopard look like a cheetah? Like 2 8 hours agopermalink Franz Weber This is called cost reduction. They will promise a tiger, and all you get is a little kitten. Like 6 7 hours agopermalink Rensol So the google pixel 7 will have 1Mp camera only? And the rest will be interpolated/recreated. Like 4 10 hours agopermalink Jefftan is this tech in pixel phone digital zoom already Like 0 11 hours agopermalink KodaChrome25 From their paper: "We also observed the model to generate very continuous skin texture in face super-resolution, dropping moles, pimples and piercings found in the reference." Look! It even does retouching. Like 3 11 hours agopermalink wcan It would be more insightful to compare an actual high res image against a these techniques performed on the same image whose resolution was reduced. Like 5 11 hours ago*permalink dr.noise the reference high-res image is literally on this page. Like 10 11 hours agopermalink Baz T I think he means to take an actual high res image of the subject, then compare it against the "enhanced" image... in short, to see how accurate the results are. Or is that what is being shown by the last figure? I don't know Like 2 10 hours ago*permalink wcan Yes, it does say "reference", but the article doesn't make that clear. Like 1 10 hours agopermalink il_alexk @wcan and @Baz T. That's what the word "reference" is used in image processing. Now it is clear, right? Like 0 9 hours agopermalink Hasa To me the only valid test is taking a low-rez image with e.g. a camera like my old Canon powershot 3.2 Megapixel and compare to an image of the same, format-filling subject taken with a 40 + Mpixel sensor. Also there needs to be not easily recognizable subject matter. If a face is recognized as "a face" by the algorithm it is easier to uprez the face with a sharp eye than say the eye of a bird half-hidden behind branches, possibly not recognized by the algorithm as a bird. A much more intelligent algorithm is needed in that case to recognize the eye and also to provide feather detail. Like 0 3 hours agopermalink Brian F Flint This is worrying. Look at the teeth of the two people. Google is adding detail which never existed in the original low resolution image. It looks good but it is false enhancement according to this example. Like 7 11 hours agopermalink ottonis Exactly! AI can enhance signal but it can only "imagine" missing information and replace it by something invented. THis invented information may come very close to the original, but it may also miss it by a mile. Like 7 10 hours agopermalink AndreyBess At the same time, I can make a photo in a way so that no "missing" information on the photo. So, who knows what is the truth. Like 0 4 hours agopermalink Hasa Well, the worrying thing is that in the future Google will compute "I recognize Mr. Flint and according to his dental records he should now be missing a front tooth, so I will use teeth from a different image minus the missing tooth" LOL Like 0 3 hours agopermalink ottonis Does it improve signal? Yes! Does it enhance and upscale images impeccably? Definitely! Can it completely and confidently/truthfully reconstruct missing original information? Most probably not! That's why CSI-like 100x enlargements are still not possible and most probably never will. Like 3 11 hours agopermalink AndreyBess Can a real photo be made so that nobody will recognize the model in the real life? - Yes. So, there is no way to get THE exactly image of a subject. Like 0 4 hours agopermalink Fotoni Could be useful for surveillance purposes to improve guessing when details are too blurry. If you are trying to create resolution from these tools, it is the same as asking computer to give you winning lottery numbers. It can be convincing, if you do not know the actual winning numbers, but the illusion fades instantly when you realize you won nothing. Like 4 12 hours agopermalink brn It would be terrible for surveillance photos. It creates detail that may not be what was really there. This tech is useful for creating better looking photos, as long as they don't need to be representative of what was actually there. Like 1 48 min agopermalink Graham Meale Well. Might as well ditch all my current gear and dust off my 3-megapixel point-and-shoot from twenty years ago. Like 7 12 hours agopermalink MichailK when hardware gets fast enough for real time video , this how we shall have 16K TV and nobody would be able to say what is real and what is computed on their regular 55' TV Like 0 13 hours agopermalink Sergeg I love my Super Zooms with their 1/2.3" sensors: LUMIX FZ300, FZ80, Nikon B700 and Canon SX50HS (now sold). The small diameter of the Circle of Confusion, allows for their 24X and 60X optical zoom ranges respectively. I use them mainly for my 4K Video work with excellent results. Their low pixel count restricts their use for serious stills work other than online image posting and social media. However noise issues aside in low light, they are capable of excellent results, particularly the FZ300 with it's constant f2.8 600mm equivalent optics. The SX50HS had a slower 50x zoom which produced amazing images, especially in RAW. So if this Ai Enhancement could be incorporated in camera at the sensor upstream of the capture stage, then the small sensor format could have a quantum leap in image quality aquisition with the amazing optical zooms which are cheaper to manufacture. To some degree, DJI are doing this on their Air 2(s) drone range. Just thinking out load, thoughts? Like 4 13 hours agopermalink ParametricPi One of my best photos ever was taken with a travelzoom compact camera by Lumix. Not practical in low light, but very fine in bright conditions. I am a happy Gcam user for less noise and more detail in smartphone pics and have always thought that I would love to be able to load Gcam onto a compact camera, or have Gcam operate my travelzoom from my phone. The lens on that little camera is fantastic. Like 1 12 hours agopermalink apsc4ever A little revolution probably will take place when the compacts get serious with computational stuff. Scary to then not directly see the difference with an unhanced FF shot. Still i would go for the 'uncorrupted' pipe line: good lens, fine apsc or FF sensor and the regular processing engine. OVF guy. And have a razor sharp SX50HS backup. Like 1 10 hours ago*permalink ParametricPi Define uncorrupted. There is loads of processing in a jpeg that we call "straight out of camera" and every serious photographer does edit. Some a bit, some a lot. An image created with computational photography can be for instance a stack of images to reduce noise, nothing suspect about that. It can be HDR, nothing suspect about that. It can be correcting lens abberations of various kinds, nothing suspect about that. Where do you draw the line in terms of corrupted? I have played around extensively with RAW files exported from Gcam. They are low-noise, fairly flat and realistic images, squeezing as much from a tiny lens and tiny sensor as they can. The weird and fake smartphone look is more overdone editing than the computational stuff being weird. Mind that a better camera lens is closer to the mathematical model of the theoretical ideal lens. A computational RAW file with 100% dynamic range/no clipping, depth information for perfect DOF calculations, no noise etc is pretty exciting. Like 1 7 hours agopermalink apsc4ever Sure thats why the ' ' around the word corrupted. No news here. Just hate the wrongv oversharpening, awful hdr , not nice fake bokeh etc eyc etc. Besides that already stated the computational can get very close to a 'clean' FF shot probably . I would be happy with the already regular pipeline when all components are very high quality. Then no need for loads of computstional stuff in my book. Photography was utterly simple in the film days (maybe the chemical process is a bit complex) . But its just capturing light. I would say keep it pure and Ai can F##k off Like 1 5 hours ago*permalink dr.noise @ParametricPi I think using "computational" towards JPEG or HDR is not right. Of course there are ALWAYS some computations in digital photography, but the term "computational" appeared not long ago and specifically implies something really fake, like fake lighting, fake skin/eyes/ age, fake zoom or fake DOF. Not just JPEG processing. Like 0 1 hour agopermalink Richard Costin Wow, such a wide gamut of whiney condescension on here. No it's not perfect but it's pretty incredible and only going to get better on the coming years. Some examples of this going from say an 8mp image to a 50mp one would be interesting. I imagine the final result would much more closely match the reference in those instances as there is more to reference in the starting point. Like 7 14 hours agopermalink ZapperVT The examples they give are going from 64x64 to 256x256 (an increase in 16x), and then up to 1024x1024 (an increase in 256x). 8Mp to 50Mp is only a 6.25x increase, so the results would appear much better. But as for "only going to get better"; this technique is CREATING detail that isn't there in the original low-res image and using AI to make that created detail look plausible. Improving AI will make the created detail look more realisitic, but it will always be created detail. Such created detail may differ substantially from the actual subject and no technology will change that limitation. Like 2 9 hours ago*permalink spider-mario That cheeseburger looks disgusting. Like 1 15 hours agopermalink Vermeero Better not upscale that one Like 1 11 hours agopermalink Vermeero The first picture in this article shows a 64x64 pixels square, a 256x256 and at last a 1024x1024 square, which are all differently scaled in their presentation. Each one must be 4 times wider than the previous one, while they're being displayed two times wider. So to me it is not a good and true example. It says 4x between each square, which is true regarding the surface area... but the mentioned amounts of pixels make no sense. Like 4 15 hours ago*permalink Christian Unger The world clearly has come to an end, if the only reference you can come up with for "Enhance" is "CSI". What a shame. Like 2 16 hours agopermalink crisdesign This technology looks amazing, and it's incredible the pace of progress. Can't wait to sew it used in real applications may be for subtle scenarios suchcas enhancing corners of wide-anglelenses, or eliminating field curvature to ensure uniform sharpness edge to edge. Like 3 16 hours agopermalink RachelCooper Pretty impressive despite the limitations. For restoration jobs or simply to update those old 2mp images lurking on ones backup; nice to have tool - who knows maybe in a few years we can right click on a low res jpg in a folder and select "upscale to 25mp"... Like 0 17 hours agopermalink AsterAstraAsteria Using this for restoration is dangerous... comparer the Reference column with the SR3 column, the SR3 restoration looks real but has significant deviation from the ground truth. For restoration work we want the result not to *look* real, but to *be* real. Inventing details is not acceptable. Like 9 16 hours agopermalink Michael Berg @AsterAstraAsteria what would a conservator do when restoring a missing piece of a painting? Like 1 14 hours agopermalink ZapperVT @AsterAstraAsteria: There are a lot of applications where looking real is sufficient. Photographs serve other purposes other than documentation, and not every photograph depicts reality. Like 1 9 hours agopermalink Franz Weber This is the end of tele lenses. Photographers can shoot the Olympics with a ultra wide angle lens like 15-35mm or a mediocre 24-105 STM. Then they can shamelessly crop afterwards to 4.500% and upres the result to match a 600mm 2.8 L lens. Like 3 17 hours agopermalink Michael Berg So ... good news then? Like 1 14 hours agopermalink sglwck Its actually worth to go to their blog and click links, e.g. this with more examples: https://iterative-refinement.github.io. In case of "nature" pics they're "class conditional" which means, I suppose, algorithm can only scale photos with specific content. And in the link above we see that some, e.g. foxes, are scales, erm, less impressively than faces. Nice gimmick, but it is still just neural net with all its limitations. Like 1 18 hours agopermalink dr.noise Yes, basically it's a trained painter who can paint a bigger picture based on a small one Like 0 53 min agopermalink Jeff Luo Maybe the government had this technology long time ago. Like 1 18 hours agopermalink Kondamin They could have had it but it would have been very very slow Like 1 17 hours agopermalink The Mad Kiwi Yeah, but clearly it has surveillance applications. Like 1 13 hours agopermalink dash2k8 Does this mean one day megapixels won't matter anymore? Like 0 18 hours agopermalink piratejabez The future is now ;) Like 0 18 hours agopermalink Franz Weber Megapixel will always be important in order to gain respect from fellow photographers. If you carry a Phase One or Fuji medium format camera you will be the man no matter what. Like 6 17 hours agopermalink ParametricPi @Franz, don't forget a very long, thick lens on that medium format camera. Like 2 14 hours agopermalink Rodrigo Pasiani And lens hood! Like 2 9 hours agopermalink tolleknolle Oh well, time to start saving for iPhone 14 Like 0 20 hours agopermalink offtheback Save your money.Get a refurb Iphone6 Like 4 20 hours agopermalink Senior Software Developer Wake me when it's on GitHub. Like 1 20 hours agopermalink the-genuine-particle Wow, this is so cool, never thought I'd live to see the day that Super Troopers would become a true story. Very cool stuff. *Enhance* *Enhance* *Enhaaaance* Like 0 20 hours agopermalink DimLS The final image is a mean of general data analysis. It's a probability game. There are possibilities John, low res photo to get characteristics of George and Lukas and make a new John. Invented details are something new and sometimes not possibly actual may be (the black dots of the leopard). Interesting tech, but virtual. Like 0 22 hours agopermalink Discombobulated fake pixels. Like 1 22 hours agopermalink RedFox88 Most likely their examples are starting with the higher resolution and downsize to the smaller ones. Like 0 22 hours agopermalink zakk9 As others have pointed out, it's clearly inventing stuff. If this product is sold to authorities that would use it to identify people, it's becoming dangerous. On the man in the bottom example, key characteristics around his eye have been changed, even the size of the pupil. For photographers, there are already very good products available, like Topas Gigapixel AI, which actually makes things look like the original photo, only at higher resolution. In the headline of the Google blog, they use the words "High Fidelity Image". High fidelity is not what this is, and that headline looks like an attempt to mislead readers into believing that it is something that it isn't. Like 17 23 hours agopermalink WJMWJM Fidel, like Castro. Or castrated fidelity. Like 0 23 hours agopermalink Otaraka Identikit and eye witness isnt exactly perfect either. Obviously misuse and incorrect understandings are a concern, but there might be some room for being helpful as well. Like 0 21 hours agopermalink pdelux "If this product is sold to authorities that would use it to identify people, it's becoming dangerous" Any different to a sketch artist, or witness description of a suspect? Like 2 19 hours agopermalink badi ""If this product is sold to authorities" - well, first let's assume that whatever authorities are not as dumb as to buy a tech based on an internet headline. And, also, the important factor is percentage of success. If any algorithm (be it inventing details or not) manages to do a 99.9% successful identification from billions of use cases, it is good enough for the job. It is not like a robot will come and sentence you to life prison without human intervention. It is just a potential identification, that could give a huge head start, especially, in urgent matters/ investigations. P.S. I totally hate and I am 100% against the idea of public monitoring and continuous identification of people as they wander around the town .... but unfortunately rather sooner then later i guess we'll end up there. Like 0 15 hours agopermalink Michael Berg @zakk9 You are thinking in absolutes when you reject the term "fidelity". I suppose this is correct but I think it is more useful to think of this technology in relative terms. As such it is "high fidelity relative to to competing technologies or methodology". Like 0 14 hours agopermalink apsc4ever @pdelux "Any different to a sketch artist, or witness description of a suspect?" Yes! Sketched witness descriptions are not made 24/7 from millions of people passing trough a random street and stored for ever by an untrustworthy conglomerate funded to do work for a government that snuffs out dissidents. Like 0 10 hours agopermalink badi Well actually in case of police identification, it is exactly as it works, as in state of the art, actually. Also previously it was exactly the same - but more crude, more "brute-force" - with the catalogues of noses and eyes :) https://www.bbc.co.uk/ideas/videos/ do-police-sketches-actually-help-catch-criminals/p08dllt4 Like 0 9 hours agopermalink twomonger For making images for artistic purposes, this is fine, the computer is simply filling in areas with artificial content of what it thinks it should be there. For journalism or forensic purposes, this is extremely dangerous and should never be used. Uneducated people or those who are easily mislead will literally think of this as the "enhance" of CSI. It is possible some may start using it to "upscale" faces or license plates, which is extremely worrisome. It doesn't help that many of the companies promoting this technology don't make it clear that the content is entirely artificial and constructed, with NO reflection on reality. Even this article is misleading - you're not "extracting" any new information, simply filling in the gaps with fake artificial content from whatever the algorithm thinks should be there. I hoped the editors here would have a deeper technical understanding than this. Like 15 23 hours ago*permalink UrbanHobbit The face examples are interesting at a sort of technophilosophical level - if I understand the mechanism, missing details for one face are filled in with minute contributions from millions of other faces. Photos of the world used to construct more photos of the world. Although this makes me wonder about the contaminating effect of allowing Super Resolution results to become part of the detail-filling dataset... Like 1 23 hours agopermalink badi " the contaminating effect of allowing Super Resolution results to become part of the detail-filling dataset" - very interesting. I guess if this thing goes .. live let's say, in one way or another, this would be the case. For example applying a subtle improvement layer on pixel phone portraits. This will "standardize" a tiny bit billions of portraits that can go into this dataset. And in time all portraits will look the same - sort of back to stick figures :)) Like 0 15 hours agopermalink ParametricPi No, it doesn't fill in with minute contributions of millions of faces, at least not to my understanding of the text. The self-learning algorithm compares its own upscales/noise reduction with originals and learns to improve itself. It gets better by making better choices on how to upscale/denoise certain combinations of neighbouring pixels etc. From what I understand: we always call noise random, but it is random around the original base value. So there is information in noise, even if a noisey pixel is not a perfect measurement of the light that fell on that pixel. Apparently the creation of noise from the low-resolution photo allows better statistical analysis. Like 0 12 hours agopermalink badi In the google blog it explains better - basically as you said, but using trained NN to do it. Anyway the results themselves hint that this is the case, as the final images show fine details rendered that is clearly not present in any way in the base low-res data. Sure, no other data will "contaminate" or get into the already trained data. But it is an exercise of thought. As I was saying - if this gets into generic cameras or photo PP apps, slowly the images will also get into various learning sets. Like 0 12 hours ago*permalink stevevuoso Let's try it out on the Bigfoot photo. Like 12 23 hours agopermalink Welsh Look honey - it's a UFO! Run inside and bring out our worst camera! Like 0 2 hours agopermalink ovatab could be another example of swapping before and after pictures now it's done by smart AI and requires all the usual stuff like trawling socials, clouds, neural blockchains and crypto subscriptions and soon we will get more excited looking at detailed portraits blown up from 4x4 pixels originals Like 1 1 day agopermalink HDPro Brasil Not impressed, the Jaguar face is clearly invented and has little to do with the actual photo. If you want a high resolution photo, use a high resolution camera or smartphone, both are getting better as we speak. I showed today a Canon 5DM4 user a photo of my dog, taken with the Sigma DP2 Merrill from 2012 and he was impressed. Like 1 1 day agopermalink Michael Berg It's very good for photos where a heavy crop was useful. Especially on mobile phones that don't have very good optical zoom - even the best of them. Like 0 14 hours agopermalink Tonny B Good Sorry, I doubt we were shown a genuine results. I assume the algorithm 'knows' how leopard and a human supposed to look. No problem here. BUT look at the leaning branch between leopard's legs... It's not even present /visible in low-res (which is supposed to be a source for synthesized image). We only have slightest pixel variations, which can be interpreted a million different ways or at least shapes. And then suddenly we get a detail perfectly true to the reference! How likely it is? Any ideas? Like 1 1 day agopermalink thx1138 Well the branch is present if you look closely and the fact the super resolution has filled in the poorly visible detail is impressive. Also who in the real world as a photographer would be using such a ridiculously low image as a starting point. What I want to see is how it upscales say a heavily cropped image of say a bird, that still has a few MP but only useful for a small print, and if it can produced an image with 2x the number of pixels and that you cannot dsitinguish it from a native image from a higher res camera. Already Gigapixel AI can do this. I can upscale an A9 image to A7RIV resolution and not be able to tell the difference. But this only works if the source material is good quality ie excellent feather detail etc. Like 1 23 hours agopermalink badi another interesting application (i think): Provide the NN with some images of the subject. And then, with it's training and the subject at hand samples, fix images in both resolution, focus, details, etc. For example: you shoot a whole day for some birds, etc. Select a few images of that specific bird/animal/human with high details. And then the NN could "fix" any image, be it blurred, miss-focused, heavily cropped, etc from that sequence, being *first* faithful to the master samples. Like 1 15 hours agopermalink Tonny B Good badi ...Or for people photos. Show the machine how person's face is supposed to look on a good day, and let an AI fix the whole series! Wohaaa! Done! ))) Like 0 4 hours agopermalink Petercatt So show us actual photos of the same models at matching high resolution, so that we can compare them to the enhanced images and see where fantasy details have been added. Like 7 1 day agopermalink Upritbass It's the last array of images in the article. Like 4 21 hours agopermalink doctor digi Clever. But basically it is inventing detail based on thousands of photos of people / animals. You aren't going to be able to resolve number plates or text like this. I'm afraid Hollywood wins out in those cases :-D Like 0 1 day agopermalink s1oth1ovechunk If it was trained on number plates or text you could probably resolve them much better than a human looking at the blurry pixels, which is exactly the point of this kind of 'enhance' operation. They aren't 'inventing' anything, they are inferring details based on the blurry input. Like 5 1 day agopermalink JakeJY Actually such methods can resolve details like that far better than a human can. There's a few license plate image enhancement projects that aim to do just that. It's not going to be 100% accurate, but it'll give you possible candidate answers that are far better than a human guessing (especially if there is video footage so multiple frames can be analyzed). Like 0 1 day agopermalink thx1138 OMG, it's amazing people's misunderstanding of computational photography. BTW doctor gigi, do you think the Sony A1 and Canon R5/R6 for example have fake AF because they use deep learning to allow animal eye AF. Like 1 23 hours agopermalink spider-mario Arguably, the situation with AF is different, in that the role of AF is merely to decide how far to focus, and if it doesn't get it right then it can be seen quite clearly even without a reference: the subject is blurry. In contrast (no pun intended), here, there is a relatively high risk of output looking accurate without being so, which might give more confidence in it than is warranted. Of course, for some use cases, it may be perfectly acceptable (e.g. if I simply want a plausible photo of a tiger), but in others, it may have to be evaluated with more scrutiny. Like 0 12 hours agopermalink mr.izo this upscaling is possible only via "ia", that comes from huge collection of the images of people and animal faces etc. now, where amd most importantly how did google get and use this photo database, ia another very very important question. Like 0 1 day agopermalink s1oth1ovechunk That's an easy question to answer. Look at the paper. It looks like they get the training/testing from imagenet: https:// www.image-net.org/ Like 0 1 day agopermalink Inspired Shot Perfect for facial recognition and population tracking. Like 1 1 day agopermalink Michael Berg If you mean that in a negative way, consider the benefits too. If you are assaulted somewhere, or your car is stolen, or if you are the victim in any other crimes that take place in public spaces, a camera might be around to identify the responsible person, bring them to justice and recover your property. Not saying for or against, just that these kinds of arguments are often one-sided in their criticism. Like 0 14 hours ago*permalink YuryVilin meanwhile, trad. camera makers just keep releasing increasingly larger and more expensive lenses... Like 0 1 day ago*permalink Total: 90, showing: 1 - 50 << First < Previous 1 2 Next > Last >> You may also like HBO's 'Hard Knocks' 5G-enabled drone Video: FPV Turn any room features incredible captures drone takes a into a FPV drone footage elaborate footage deep dive down right-side-up of the Dallas from the MLB's world's camera obscura Cowboys' stadium 'Field of Dreams' tallest with the Aug 26, 2021 game waterfall BonfotonUP kit Aug 19, 2021 Aug 17, 2021 Aug 12, 2021 Latest sample galleries Rokinon / Samyang AF 24mm F1.8 FE sample gallery Nikon Z 14-24mm F2.8 S sample gallery (DPReview TV) Nikon Z fc sample gallery Canon RF 28-70mm F2 L sample gallery (DPReview TV) See more galleries >> Latest in-depth reviews 78 In-depth tripod review: Leofoto 'Summit' LM-364C reviewAug 30, 2021 at 17:52 Leofoto is one of the best-known names in the world of affordable tripods and supports, but not necessarily for the best of reasons. With the LM-364C, Leofoto appears to offer some innovation and high value - so how does it perform? 2033 Nikon Z fc Review reviewAug 26, 2021 at 15:00 The Nikon Z fc is a midrange ILC with the same eye-catching design as the company's classic film cameras. It's not just a good-looking camera, though: it's quite capable. Read our review to find out if the Z fc is right for you. 103 In-depth tripod review: ProMediaGear Pro-Stix TR344 reviewAug 17, 2021 at 13:00 ProMediaGear is well-known within the photography industry for their long-lens gimbals and flash brackets , but with the Pro-Stix TR344, this family-run company is making a play for the premium 'systematic' tripod market. Read our in-depth review to find out how it performs. 263 Tested: Sony's trio of compact 24mm, 40mm and 50mm G prime lenses Aug 16, 2021 at 13:00 Earlier this year Sony announced a set of three compact full-frame E-mount prime lenses: the 24mm F2.8 G, the 40mm F2.5 G, and the 50mm F2.5 G. We've been taking a detailed look at their optical performance - click through to read more. 121 Panasonic Leica 25-50mm F1.7 ASPH field review reviewAug 12, 2021 at 16:02 Panasonic's Leica DG Vario-Summilux 25-50mm F1.7 ASPH is a fast zoom lens aimed at photographers and filmmakers using Micro Four Thirds camera bodies. Read our review to find out how it performs. Read more reviews >> Latest buying guides Best cameras around $2000 in 2021 Aug 11, 2021 at 23:39 What's the best camera for around $2000? These capable cameras should be solid and well-built, have both speed and focus for capturing fast action and offer professional-level image quality. In this buying guide we've rounded up all the current interchangeable lens cameras costing around $2000 and recommended the best. Best video cameras for photographers in 2021 Aug 11, 2021 at 21:16 Most modern cameras will shoot video to one degree or another, but these are the ones we'd look at if you plan to shoot some video alongside your photos. We've chosen cameras that can take great photos and make it easy to get great looking video, rather than being the ones you'd choose as a committed videographer. Best cameras for Instagram in 2021 Aug 5, 2021 at 00:41 Although a lot of people only upload images to Instagram from their smartphones, the app is much more than just a mobile photography platform. In this guide we've chosen a selection of cameras that make it easy to shoot compelling lifestyle images, ideal for sharing on social media. Best drones in 2021 Jul 24, 2021 at 00:43 If you're looking for the perfect drone for yourself, or to gift someone special, we've gone through all of the options and selected our favorites. Best cameras for vlogging in 2021 Jun 16, 2021 at 23:54 Whether you're just sharing clips with friends or you're launching an online on-camera career, vlogging matters. We looked at cameras with selfie-friendly screens, wide-angle lenses, microphone inputs and great video quality, and selected the best. Check out more buying guides >> Featured Videos Camera Shoulder Bags compared: Peak Design, Lowepro, Vanguard, Think Tank, Manfrotto & Wotancraft Camera Shoulder Bags compared: Peak Design, Lowepro, Vanguard, Think Tank, Manfrotto & Wotancraft Nikon Z 14-24mm F2.8 S Review - The Nikkor F2.8 trio is complete! Nikon Z 14-24mm F2.8 S Review - The Nikkor F2.8 trio is complete! Why lower resolution sensors ARE NOT better in low light Why lower resolution sensors ARE NOT better in low light Finished challenges Velvet Boat on Loch Breasted Insh, Scotland Canon lens 400mm 5.6 Gone Fission Roller Velvet Boat on Loch Canon lens 400mm 5.6 Gone Fission Breasted Insh, Scotland by itamarcampos by Rodger in Roller by by BRTphoto from Favorite Camera Edmonton Rhinosaur from Minimalist Gear - Specific from Flying from Just a Photography Equipment Challenge Without jiffy Engines Discover more challenges >> Latest articles RED teases new cinema camera that appears to record 8K video at 120fps Jarred Land, CEO of RED Digital Cinema, shared an Instagram post that shows off the screen of an unreleased RED camera system. Aug 31, 202127video news Polaroid's new Now+ instant camera uses your smartphone to unlock specialized capture modes The Polaroid Now+ offers a collection of new and improved shooting modes, but you'll first need to connect it to the Polaroid mobile app via your smartphone. Aug 31, 20218camera news Video: A look at Canon's 40-year-old AF lens, the FD 35-70mm F4 AF In 1981, Canon introduced its first and only autofocus lens for the FD system, the 35-70mm F4 AF. It's a fascinating lens, and YouTuber Kai W has gone hands-on with it on a Sony A1 to see if the archaic AF lens still works in 2021. Aug 31, 202141lens news Panasonic Lumix DC-GH5 II added to studio test scene We've shot and processed our standard studio test scene for Panasonic's Lumix DC-GH5 II (GH5M2). Our review will focus mainly on video, but talk of improved dynamic range made us want to check whether we could see an improvement over its predecessor. Aug 31, 202115 Iconic 1914 Ur-Leica camera recreated as LEGO could become an actual set LEGO Ideas user Pandis Pandus has created a LEGO Ur-Leica camera. The camera, released in 1914, could become an actual LEGO set if it gets enough support on LEGO Ideas. The Ur-Leica was the first camera to use the now-standard 35mm format. Aug 30, 202161 In-depth tripod review: Leofoto 'Summit' LM-364C Leofoto is one of the best-known names in the world of affordable tripods and supports, but not necessarily for the best of reasons. With the LM-364C, Leofoto appears to offer some innovation and high value - so how does it perform? Aug 30, 202178review Slideshow: Winning photographs from the inaugural 'Captured at Calke' photo competition A slideshow of the winning images of International Garden Photographer of the Year's inaugural 'Captured at Calke' photo competition. Aug 30, 202125 Enhance! Google researchers detail new method for upscaling low-resolution images with impressive results Google might just be turning CSI's laughable 'Enhance!' scenes into a reality. Aug 30, 2021241 ASUS Zephyrus G14 (2021) Review: The M1 MacBook Pro killer The 2021 ASUS Zephyrus G14 offers a phenomenal combination of performance, power efficiency, and portability. With its AMD Ryzen CPU, RTX 3060 GPU, and color-accurate WQHD display, this little laptop might just be an M1 MacBook Pro killer. Aug 29, 2021369 DPReview TV classics: Best (and worst) ways to clean camera lenses Chris and Jordan are taking a well-deserved break this week, so we're re-posting one of our favorite episodes from the archives. In this humorous video they discuss the right way - and the wrong way - to clean a lens. Beware of broken glass! Aug 28, 2021103video Zeiss will raise the prices of some lenses in the US on October 1 A graphic shared by Leica Rumors shows Zeiss will be increasing the price of some of its lenses at the start of October. Aug 27, 2021220lens news Venus Optics adds new mount options for seven of its most popular Laowa lenses Seven existing Laowa lenses get new mount options, including the 24mm Probe lens which now comes fitted for Nikon Z and Canon RF, while the wide Cine lenses come to Nikon Z and the L mount Aug 27, 202135lens news Prolycht's Orion 300 FS is a relatively affordable six-color LED studio light Six-color LED lights not only produce more accurate skin tones across a range of white balance settings but also create a much wider spectrum of color effects too. Such lights used to be only for those with Hollywood budgets, but Prolycht has just brought the price down. We all could be using them soon. Aug 27, 202132 Thunderbolt Drives compared: Samsung X5 vs. Sabrent Rocket 4 Plus If you're looking for fast portable external drives, and your PC or Mac has the right port, then Thunderbolt is the way to go. We compared the integrated Samsung X5 against the enclosure-style Sabrent Rocket 4 Plus. See which one came out on top. Aug 27, 202155accessory review HBO's 'Hard Knocks' features incredible FPV drone footage of the Dallas Cowboys' stadium Episode 3 HBO's series "Hard Knocks' features incredible FPV drone footage of the Dallas Cowboys' football stadium Aug 26, 202138 ON1 teases back-up and restore features coming soon to Photo RAW ON1 has been teasing upcoming features for its Photo RAW software over the last week. The latest new feature is a back-up and restore function that will allow users to save and restore all their critical albums, presets and other settings within Photo RAW. Aug 26, 202119 Nikon Z fc Review The Nikon Z fc is a midrange ILC with the same eye-catching design as the company's classic film cameras. It's not just a good-looking camera, though: it's quite capable. Read our review to find out if the Z fc is right for you. Aug 26, 20212033review Sigma's Optimization Pro software now has native support for Apple Silicon Sigma Optimization Pro version 1.7.0 brings, amongst a few smaller updates, support for Apple's M1 computers. Aug 26, 202114 Tamron reveals $699 18-300mm F3.5-6.3 APS-C lens for Sony E mount cameras, X-mount version due out 'this fall' The lens features a 27-450mm full-frame equivalent focal length, weighs just 621g (21.9oz) and is set to ship later this month. A Fujifilm X-mount version is expected to ship later this year. Aug 26, 202174 Godox announces V860 III speedlight with modeling light, improved battery and more Godox has announced the V860 III speedlight (also available as the Flashpoint Zoom Li-on III R2 TTL through Adorama). It includes a modeling light, upgraded battery and more. Aug 25, 202160 Samsung disables the cameras on its new $1,800 Galaxy Z Fold 3 smartphone if you unlock the bootloader One Android expert compared the move to something Apple would do, saying 'It's really obnoxious of Samsung to do this, and I am in two minds now about keeping the device and working on a fix for the issue, or simply returning the unit in disgust.' Aug 25, 2021189 These custom-built hard cases protect camera rigs for Canon, Fujifilm, Nikon, Sony and others Italian case company HPRC has introduced a new line of cases custom-designed to hold specific camera equipment from a wide range of manufacturers, including Atomos, Blackmagic, Canon, DJI, Fujifilm, Leica, Nikon, Panasonic, RED, Sony and more. Aug 25, 202184 Hands-on with the Canon RF 14-35mm F4 L IS USM Canon's RF 14-35mm F4 L IS USM is billed as a more affordable, lightweight alternative to the flagship wide-angle zoom for RF mount, the RF 15-35mm F2.8. Click through this article for a closer look at its features, specifications and handling qualities. Aug 25, 2021380 Slideshow: Winners and finalists of AAP Magazine's #19 SHAPES competition Here are the winning and finalist images from All About Photo (AAP) Magazine's #19 SHAPES competition. Aug 24, 202110 Production issues, lack of funds have brought the Reflex SLR project to a halt In 2017, the Reflex SLR project was successfully funded on Kickstarter. The project promised a new, modular 35mm camera, but it seems the endeavor has run out of funds throughout the development process. Aug 24, 2021193 Video: Curiosity rover captures 360-degree panorama of Mount Sharp on Mars, showing changing landscape NASA's Curiosity rover is in its ninth year on Mars, and the rover continues to do important scientific work. The rover recently captured a 360-degree panorama of the Mount Sharp area, showing a fascinating landscape and discovering more about the red planet's history. Aug 24, 202190 You will soon be able to merge HDR, panoramic photographs in Capture One Capture One has announced it will be adding HDR and panorama merging--two of its most-requested features--to a future version of Capture One 22 that's due out later this year. Aug 24, 2021173 Updated: Pentax K-3 Mark III sample gallery As we work towards our full review of the Pentax K-3 Mark III in the next few days, we've updated our previously-published gallery of sample images, including both out-of-camera JPEGs and converted Raw files. Click through to take a look! Aug 24, 2021133sample gallery The UK's largest news agency announces partnership with Sony The UK's largest news agency, PA, has partnered with Sony to provide its 90+ journalists with Sony cameras and lenses. Aug 23, 2021672 The CP800 is a modular, all-in-one film processing machine from FilmNeverDie The Compact Processor 800 is a modular film processing machine that can develop up to eight rolls of 35mm film or four rolls of 120 film at a time and pulls chemistry from integrated tanks. Aug 23, 202193 Tips or suggestions? Contact us! More articles >> www.dpreview.com Follow us Mobile site About * About us * Work for us * Advertise with us * FAQ * Feedback / Contact us * Interest-Based Ads * Privacy * Legal Editorial content * News * Camera reviews * Lens reviews * Printer reviews * Buying guides * Sample images * Videos * Editorial enquiries Cameras & Lenses * Cameras * Lenses * Camera search * Camera comparison * Lens search * Product timeline * Browse all products Community * Community Guidelines * Forums * Challenges * Galleries * My Profile * My Settings * My GearList All content, design, and layout are Copyright (c) 1998 - 2021 Digital Photography Review All Rights Reserved. Reproduction in whole or part in any form or medium without specific written permission is prohibited.