Sponsored: Sign up at nordvpn.com/macroking to get 4 extra months on a 2-year plan!

Those AI generations are basically a legitimate, modern-day 5th panel to the comic lol.

The issue with Yandex yielding wildly incorrect results at 30:19 is because the URL was pasted in the "normal text" search box - it will be treated as a garbled mess of characters which the search engine isn't tuned for. Instead, you have to press the camera icon, on the right side, which should appear if you are already in the "Images" tab; under the "Drag an image here" box that pops up should be another text box specifically made for image URLs - it should retrieve the image, and plug it into the image search engine which seems to be an almost independent thing compared to the text search. Note that you have to press the grey search button beside the box, not the large yellow one on top. I'm sure that it might not be significantly faster than dragging the image itself, especially with macros and what not. Anyhow, I suspect that this quirk is what had distinguished Yandex's image search from the rest since whenever I tried Google's reverse image search I'd notice that they would accompany the query with some relevant text which might be them using some form of automated computer vision, and labelling, to contort an image search into their primarily text-based paradigm. Nevertheless, the aforementioned oddity seems to have stopped quite recently; perhaps they finally (properly) integrated Google Lens into their image search.

man, how you strive for technological excellence even with something as trivial as including a screenshot in a video really resonates with me :)

Sorry for all the comments, one final thing about Waifu2x and that halo effect you get on text. It's from the model you're using. Switch to the "2-D illust (CUnet Model)" for almost everything (always for logos and screenshots) and you will get none of that stuff whatsoever. I just tested it on an image with "2-D illust (UpResNet10)" and I got the halos, switched to the CUnet Model and no halos at all. It's the BEST model by far!

A way I found how to remove the sharpening white halo effect around text is to: 1. Create a duplicate above the original image. 2. Use: Filter > Other > High Pass on the duplicated image. (The radius of the High Pass has to be fine tuned to get the best results) 3. Inverting the resulting gray image. 4. Setting the bend mode to Hard Light and reducing the Opacity/Fill to just about where the halos disappear. This essentially creates an edge blur filter. It does soften the image a little bit, and that's why it's needed to fine tune the High Pass radius to the shortest distance that still removes the halos, but doesn't introduce too much blur.

(18:00) Worst crime I've seen is that I've seen someone upscale pixelart by nearest neighbour by a non-integer value, and then re-scaled it with bilinear.

Can we just agree that it was obviously intentional at least the last one, possibly the whole increasingly less quality, since that is the whole point of the joke? Still works as an example of course.

Im forty minutes into this and just now realising this isn't eight minutes long

KZfaq is still showing this video as 360p, the irony

1:23:00 You can use something like the Enhancer for KZfaq extension, which adds extra video controls including the ability to directly export the current frame

This is one of the most interesting/satisfying videos I've watched in a long time, yet it's about such an unconventional/niche topic, it feels tailored specifically to me. I could just like, watch you do this, for hours.

It is important to recognise that the resulting images from using models to recreate images are not screenshots. For many uses they are great, but for others they can be problematic.

After years of messing with screenshots by hand, it's truly amazing to see what a real professional does, and how efficient the workflow can become. Excellent video, I enjoyed every moment.

All the artifacts in the Windows screenshot could maybe be removed by palettizing down to 4-bit color.

mega tutorial that I will watch thoroughly and then never have a use for in my life

This is an amazing resource, I have used some of this methods over the years, but this gives me greater context on how to use then combined. Thank you so much!

Before watching this video i didn't realize how much i would like video editing. I always love obsessing over every little detail and getting things perfect and seeing this video of taran doing the same frankly just makes me happy

Regarding waifu2x split size, I always use 128 (or at least a number that's in the section in the dropdown where 128 is) because the other numbers often produce results that look like you've stepped into a mirror maze/room attraction at a circus or fairground (it's honestly hard to explain until you see it for yourself).

LCD effect you can do with Photoshop action. I'd do it for you, but i need to work, so i will try it in few days. In short: • enlarge image using Nearest Neighbor to 900% or 600% (if you want very accurate result) or to 300% if just want red, green, blue bars side to side. • separate image into 3 layers each containing each channel: r, g, b (use copy layer and Channel Mixer adjustment for each) • make 3 layers with patterns of red, green, blue subpixels Need to premake patterns: 3x3 pixels for 300% line of 2 or 3 column of filled pixels and blank surrounding; 6x6 for 600% but 2 pixel wide columns; 9x9 for 900% an oval shape with blank surrounding. • apply the of r, g, b layers to the corresponding pattern layers as clipping mask • to each layer set add as a clipping mask a Fill Layer with corresponding color and set to Darken blend mode • create a black background layer if you are at 900% scaling cause subpixels vertically do not touch, so you'd need some space between them. • merge all (or select all and create a smart object so you can tweak it later) It's not a hard thing to write in PS. Don't know as for Premiere or After Effects. Example (manual, not an action yet) imgur.com/a/pG4djuB

I tought painting restorations were cool but this is next level hahaha

The section with removing JPEG compression may be helped with a Median filter - it samples all pixels within a given radius and chooses the median, which is the most common colour. In most cases this is the intended colour without compression artifacts. The only issue is it won't preserve borders, but combined with a Maximum or Minimum filter it may be possible to clean up.

The definitive masterclass tutorial you never knew you needed... nor existed!

Regarding the amazing tunic filters: Why don't you reach out to the creators? There has to be someone talented making it and you having a chat with them is definitely some content I would love watching.

When I wanna get two images aligned as close to pixel-perfect as possible, one trick I often use is to set the top layer to Grain-Extract blending mode; that way it gets pretty easy to see where the images match and where they don't, and you can easily see when you overshoot alignment as the mismatching edges will switch colors/brightnesses. That can help even when you need to warp one image to make it match the other. It doesn't necessarily guarantees you'll be able to get 100% perfect, but IMO it makes it a lot easier to get it very close. And if for whatever reason you can't use grain-extract, a similar effect can be obtained by inverting the colors of the top layer and setting to 50% transparent.

1:22:45 In Firefox, if you double-right-click on the video there is a "Take Snapshot" option which saves an image of the current frame of the video, in the native resolution that the video is set to, with no UI on top.

Taran Van Goat

Also, if you just want to decode a JPEG that was originally a clean screenshot, there is the underwhelmingly-named jpeg2png which uses calculus of variations to make a better estimate of what the original image was (computes possible source images that could have been encoded to that JPEG).

You can get a texture of a screen fully lit (white), scale the image you want to apply the effect on so that it matches the screen texture and set the blending to multiply.

thanks for the separate ad chapter i always give vids with that a thumbs up

Wait Taran you have your own channel?!?!?!?! Always loved to see you make LMG cameos! Glad i found you! Great video my friend :) You wouldnt remember but we met once and you were suuuuper friendly. Keep grinding!

Great video, Taran! Loved every second of it.

Beautifully done

How did i watched the entire two hours video...! this is just awesome stuff! Some of the tecniques i actually use for my design work tbh, but love it!

You're awesome and so talented Taran. Thank you for sharing your knowledge with us!

this is an amazing video! you are really good at fixing the screenshot

This method has proven mighty useful for badly taken photos (and some photoshop effects + manual tinkering helps too)

55:54 something to try could be: median filter to remove text from background (content aware fill away too big objects before) and a) use blend mode darken or b) create a text-on-white layer with blend mode divide, clip the luminance range if needed and multiply on background, replacing parts like logos etc.

I love your pinned folders

to remove the white outline around the text, you could rectangle select where all the text lies, put it into another layer, then make a clean plate of that region on the screen, then put the layer with the text on top of it with the multiply blending mode, that would only leave the relative darker pixels to show

took me way too long to recognize taran, this video stands on it's own so well.

Great tutorial! Speaking of Chrome Extensions, there's a Yandex Images one that let's you right click on an image and automatically does the reverse image search in one click, pretty convenient!

You are a genius and a mad man. Respect to you, sir.

I'm not sure if you know this, but when using Curves, if you alt/option click on the triangles, it'll toggle the effects of "Show Clipping" feature

Watching this video really makes me feel less alone in the insanity that is pixel peeping I partake in. Great video, learned LOTS of new tricks to use in my never-ending image restoration adventure :^). P.S. you can use the 'Improve Tube' Chrome extension to add the ability to take a direct screenshot from a KZfaq video!! I use it all the time and it spits out a screenshot the exact same size as the source resolution you are viewing at. Once again thanks for the extremely helpful advice!!

nice one taran

wow…now i can restore old emoticons, amazing

I just watched the whole video, although I will most likely never use any of this ^^ Your presentation and explanation is always great!

1:42:35 I believe older operating systems had different icons depending on the colour palette and colour depth of the monitor, so there's multiple different sets of icons

I always look forward to your video products, what you do is really amazing🥰🥰🥰🥰🥰

"Upscayl" is probably my new favourite AI upscaler with many different models (also ESRGAN), keeps getting great updates and is open source (super easy to use as well)

im not even interested in editing whatsoever and I watched this through. Fun.

OMG I didn't even realize it was you Taran 'til I saw your face I'm glad I stumbled upon you once again

This is exactly what I need damn thanks so much

You can disable sub pixel rendering by searching "view advanced system settings", clicking on the settings button under performance, and unchecking "smooth edges of screen fonts". Also you can disable text anti aliasing in GIMP

I use RealSR for upscaling and it's magnificent for stills. You can even use SRMD for video (srmd is faster but less precise so it can work with footage). You can use esrgan or whatever too but RealSR and SRMD are the two best ones. The way I went around the non-gui in this case is set up a BAT file with specified 4x upscale (you can have more of those with different settings) and specified input folder and output folder and I just throw in the images, doubleclick the appropriate bat file and it'll output all the files I threw in the input folder - beware that these outputs are usually huge in filesize and it's best to run them through photoshop afterwards to use its better algorithms to save the final image.

Upscyl works good and it’s constantly being updated for different types of artwork, i use it general photos

note: i believe the tunic manual pages were actually printed out and then rescanned, and that's where they get the printing look. so i don't think they have an actual effect for it.

You absolute legend

This video is completely lost on my phone fyi.

to get a better result around the text on the windows image, create a copy which is gausian blurred with high value to then use as a hsl color blend layer and mask it in manually.

Now it starts to make sense when in movies the FBI zooms 1000x into a low quality video, and somehow converts it into a 4K image

15:55 black and white is also available under the "image" dropdown on the menu bar

5:40 Also so you can compare with the original to be sure you actually improved it. 28:16 Ahhhh, now all this makes sense in the context of the tech quickie videos. I never thought about having to clean up screenshots for videos. Gives me new appreciation for all the work that can go into a picture shown for a few seconds.

55:50 so what I would do is make a duplicate of the layer and draw over the text with whatever the color of the background is and make it a darkening layer. It won't effect the text or background, only the surrounding glowing pictures. You could also use content aware fill depending on the text and background complexity. In this case the resolution is too small and it would be easier to draw by hand. But since it's only to get rid of the glow it can be fairly rough. Hope this helps someone.

My favourite tool for pixel art is MS Paint. And no, not the 2007 version or later, I mean the Windows XP version. Sure it has its shortcomings, but! It has some useful shortcuts! To instantly upscale a selection to 2x, you only need to have an active selection, then do Ctrl++ to increase or Ctrl+- to decrease

Looking though these comments just one day later, and there's so much useful stuff. Honestly, i think you need to make a second part some time later, or a similar document to the one you made for the color correction video

*Obviously* you are right about nearest neighbor. As for the CRT upscaling, it depends on so many factors it's never ever going to be "right". CRTs often differed from model to model, and looked "right" only on the artist's original display. At least early on, but that's a whole other thing. So yeah... it's a lot to think about! 😄

1:22:41 if you right click twice on the video it will bring up the KZfaq context menu but then the browser context menu from which you are able to "take snapshot" which allows you to easily get a still frame from the video unobstructed by the KZfaq UI

57:45 You can use it, and other various upscale models (like 4x-UltraSharp), using Stable Diffusion web UI, from the Extras tab.

this is the best video ever

fun fact: if you use youtube with a german ip, google/youtube will ONLY sell you movies in german dubbed version. No choice, no options. But i don't need a VPN (but it works) , Amazon for example treats you like a grown up with mutliple languages. Or, you know,,, a Bluray.

55:25 the way i remove quickly and efectively haloing around text is: 1. I select every element in an image and invert the selection so what remains is enough "clean" background and put this on a new layer 2. Edit -> Content Aware Fill. So you end up with a layer that is just the backround 3. Put the newly created layer on top of the original 4. Set its blending mode on darken 5. Mask out the areas that have been unintentionally affected. 6. Profit!

KZfaq in all its brilliance showing me this video in 240p so i can really see all the fine detail being restored

For CRT and LCD simulation, you could try loading images into RetroArch with a CRT or LCD shader enabled, and take a screenshot (though RetroArch is designed for playing games in real-time, and it's hard to take a screenshot bigger than your window or screen size). These were more meant for video games on TVs and handheld screens, but might work out for PC images as well. Be sure to avoid composite shaders since most PCs were not connected through a 240p/480i composite color connection. For printing, you could look for halftone simulators. I'm less familiar with this field, but Photoshop (and Photopea) seem to have a built-in grayscale-only halftone filter. For a less aliased result, you can scale up the photo 2 or more times using bicubic or similar, run the halftone filter at a relatively high contrast (in Photopea, 46 is the highest contrast that doesn't completely remove antialiasing altogether), then scale the image back down to the original size. If you want CMYK halftoning, you'll have to split the image into CMYK, upscale and rotate the layers, and halftone them individually. Or you can find a different tool other than Photoshop. To remove a white outline (originally used as a contrast enhancement technique), there are multiple strategies: - You could try a convolution filter with a single bright pixel in the middle and a wide dim disk or Gaussian kernel around it, to spread the darkness of text into the surrounding overexposed white pixels. However this seems to work poorly, causing the text to be gray rather than black. - You can use non-contiguous selection with a threshold of 0.5, click a black pixel, deselect anything but text, feather your selection by 1-2 pixels 1-2 times (to select text and surrounding pixels), then use Curves to bring down the brightness of text and surrounding regions. The feathering process ensures that over-brightened regions surrounded by text are darkened more than less over-brightened regions next to text. If you want to take pixelated screenshots without a VM, you have to run SystemPropertiesPerformance.exe, uncheck "Smooth edges of screen fonts", and restart your browser. I verified on archived KZfaq "Me at the zoo", that both regular and bold Arial are pixelated in both Firefox and Vivaldi.

great vid

"I need a GUI man, I can't handle command line" Proceeds to do everything to avoid using the GUI by binding actions to hotkeys and setting up macros

I've actually retyped a text mode screenshot before, saved it as a text file, then opened it in a command prompt window (with the font set correctly) and taken a screenshot of that. Granted, that's not exactly a super fast method, but all available screenshots were just that degraded that it was faster. I've also sometimes fixed a blurry upscaled screenshot by downscaling it back to its original resolution. This works very well if the original has a limited palette. I definitely would have tried it for the GRRM screenshot, since it only has two colors. Granted, this was for Wikipedia, not a video.

omg its taran... well-done KZfaq algorithim *1. a good video i wanted to watch* *2. Hey taran didn't know you had a channel :)*

30:10 One can paste urls, but instead of using the search box for it, one should click the "search by image" button (that camera-shaped icon near the "Search" submit button) and paste url into the popup. But anyway, it's always important to remember that (pretty much) all restoration is destruction, so for archiving purposes it's important to save the version that underwent the minimal amount of sequential edits, because most instances of restoration introduce additional bits of information and addition of information without information how to perfectly rollback it damages the object in general case. Also, AIs can add artifacts, so it may be better to insert not previously upscaled images into future AI models.

For image at 1:40:00 i'd try turning it into Indexed color and change the indexed colors. Also you might like Max Chroma (at YT) actions for recoloring stuff. He does magic with those actions. They are for print mostly, but you might use something from it www.youtube.com/@MaxChromaColor

I think that diffusion is better for restoring photos or illustrations rather than something so precise as a line drawing

dude i thought that said "roasting ruined screenshots" for a sec even though i watched that video multiple times... yes i actually really like these videos :D

the nearest neighbor plus other scaling option is a great way to play older games too using Magpie, i have 2/3 integer scaling to scale the game resolution up and just over my monitor resolution and downscale abit using bilinear to fit my screen

ThioJoe ThioJoe 3 weeks ago (edited) • There's an open source image filter framework called GMIC that has an effect called "CRT Sub-Pixels" that might do what you want. It's under the "Degradations" group. There's even a GMIC plugin you can install for photoshop right from the Creative Cloud desktop app's marketplace thing. It also has a crap load of other useful features. • Also Edge seems to honor Windows' ClearType setting for subpixel rendering at least. There's even a flag called "Enhance text contrast" talking about it, but it seems unnecessary to change it. • Edit: Actually it looks like you can also turn off anti-aliasing (AFTER turning off cleartype) by disabling the "smooth edges of system fonts" performance setting for Windows, which you can get to by searching start menu for "View Advanced System Settings" > Performance box > "Settings" > with 'Custom' mode uncheck the smooth edges setting. For some reason, whenever you toggle the ClearType option (whether enabling or disabling), it re-enables the "smooth edges" setting automatically, so you need to disable cleartype first. • Another Edit: Actually I just realized this works for chrome too, I just didn't realize you have to restart the browser for it to apply. 492 Reply 4 replies Mark DAGA Mark DAGA 3 weeks ago Those AI generations are basically a legitimate, modern-day 5th panel to the comic lol. 488 Reply 8 replies cURLy bOi cURLy bOi 3 weeks ago man, how you strive for technological excellence even with something as trivial as including a screenshot in a video really resonates with me :) 128 Reply 3 replies Demitri O'harr Demitri O'harr 2 weeks ago The issue with Yandex yielding wildly incorrect results at 30:19 is because the URL was pasted in the "normal text" search box - it will be treated as a garbled mess of characters which the search engine isn't tuned for. Instead, you have to press the camera icon, on the right side, which should appear if you are already in the "Images" tab; under the "Drag an image here" box that pops up should be another text box specifically made for image URLs - it should retrieve the image, and plug it into the image search engine which seems to be an almost independent thing compared to the text search. Note that you have to press the grey search button beside the box, not the large yellow one on top. I'm sure that it might not be significantly faster than dragging the image itself, especially with macros and what not. Anyhow, I suspect that this quirk is what had distinguished Yandex's image search from the rest since whenever I tried Google's reverse image search I'd notice that they would accompany the query with some relevant text which might be them using some form of automated computer vision, and labelling, to contort an image search into their primarily text-based paradigm. Nevertheless, the aforementioned oddity seems to have stopped quite recently; perhaps they finally (properly) integrated Google Lens into their image search. 110 Reply Martin Martin 2 weeks ago (edited) Two things that you didn't mention in the video (you might already know them, idk): • Modern websites use the `srcset` HTML attribute for image elements sometimes. These contain URLs to different sized images for different pixel display densities or zoom levels. So, when you are zoomed it, the browser will actually use one of the larger resolution images so it looks clearer and not having to do browser upscaling. However, at least Chrome does not do this change when you zoom in on an already loaded page. You have to zoom in AND then refresh for the larger resolution source images to be used. This can lead to better looking screenshots of websites that use images. • Another alternative to Yandex is searching for the original resolution image uploaded to the web server. A lot of the time those won't be indexed by search engines because they're not actually used on any webpages and can only be accessed by directly going to their URL. This is done to save bandwidth by both the server and the user browsing the webpage, they are dynamically downscaled AND even [JPEG] compressed by the webserver before being served to the user. You stumbled upon a lot of these (especially on Wikipedia) when you were using Yandex reverse search (see the URL has mentions about thumbnail and also a specific pixel size mention such as "500px"). You can either recognize these by yourself (easy for stuff like Wikipedia) or use a tool which knows how to parse URLs and automatically find the original file. One tool to do this is via ImgOps (I'm sure you know the website) and clicking "server - bigger version" and it will use it's predefined filters to do a search. Another tool was posted by someone else in the commons but it's a lot more advanced called Imagus (you can find their comment if you want to check it out) which has tons and tons of filters and the person even linked to a forum post with another 600 or so more filters. Fun Fact: The DOS looking text editor URL that you originally had contained "650_1200" in the URL which if you notice is actually the resolution of the image being displayed and instead the original file name was "wordstar-jdt". If you replace that resolution with "original" in the URL, it will serve you the original image from the web server which imo is actually a better image than the one you found on Yandex. This is just one example of the second point I made above. 92 Reply 8 replies Liggliluff Liggliluff 2 weeks ago (18:00) Worst crime I've seen is that I've seen someone upscale pixelart by nearest neighbour by a non-integer value, and then re-scaled it with bilinear. 36 Reply Martin Martin 2 weeks ago (edited) Sorry for all the comments, one final thing about Waifu2x and that halo effect you get on text. It's from the model you're using. Switch to the "2-D illust (CUnet Model)" for almost everything (always for logos and screenshots) and you will get none of that stuff whatsoever. I just tested it on an image with "2-D illust (UpResNet10)" and I got the halos, switched to the CUnet Model and no halos at all. It's the BEST model by far! 64 Reply 1 reply Offline Offline 2 weeks ago A way I found how to remove the sharpening white halo effect around text is to: 1. Create a duplicate above the original image. 2. Use: Filter > Other > High Pass on the duplicated image. (The radius of the High Pass has to be fine tuned to get the best results) 3. Inverting the resulting gray image. 4. Setting the bend mode to Hard Light and reducing the Opacity/Fill to just about where the halos disappear. This essentially creates an edge blur filter. It does soften the image a little bit, and that's why it's needed to fine tune the High Pass radius to the shortest distance that still removes the halos, but doesn't introduce too much blur. 26 Reply 3 replies Spicy Baguette Spicy Baguette 2 weeks ago 1:23:00 You can use something like the Enhancer for KZfaq extension, which adds extra video controls including the ability to directly export the current frame 10 Reply Sanian Creations Sanian Creations 2 weeks ago This is one of the most interesting/satisfying videos I've watched in a long time, yet it's about such an unconventional/niche topic, it feels tailored specifically to me. I could just like, watch you do this, for hours. 3 Reply maraz666 maraz666 2 weeks ago All the artifacts in the Windows screenshot could maybe be removed by palettizing down to 4-bit color. 17 Reply 1 reply Martin Martin 3 weeks ago Regarding waifu2x split size, I always use 128 (or at least a number that's in the section in the dropdown where 128 is) because the other numbers often produce results that look like you've stepped into a mirror maze/room attraction at a circus or fairground (it's honestly hard to explain until you see it for yourself). 17 Reply 1 reply Silacai Silacai 2 weeks ago The definitive masterclass tutorial you never knew you needed... nor existed! 3 Reply Hugo Spencer Hugo Spencer 2 weeks ago Regarding the amazing tunic filters: Why don't you reach out to the creators? There has to be someone talented making it and you having a chat with them is definitely some content I would love watching. 12 Reply Alek VH Alek VH 2 weeks ago (edited) For image at 1:40:00 i'd try turning it into Indexed color and change the indexed colors. Also you might like Max Chroma (at YT) actions for recoloring stuff. He does magic with those actions. They are for print mostly, but you might use something from it www.youtube.com/@MaxChromaColor

I'd try first Adobe Camera Raw Filter Denoise, i will, i'm interested now. Also for Black&White layer - the proper numbers for RYGCBM are (and should be default) 30/89/59/70/11/41. If you set it like that and put B&W layer in Luminosity blend mode, the image won't change, meaning those are the correct RGB Lightness values for primary colors.

I think using a model fine-tuned for drawing black and white stick figures would help a lot more than "vanilla" stable diffusion. Though I'm not sure if anyone trained such model yet, but it's not too hard to do by yourself...

You are a wizard !!!

RE: Photoshop Image DPI, If you uncheck "Resample", you can set the DPI to what you want, without changing the image size. I dont understand why adobe does this. Great vid! Loved your keyboard reviews too

1:33:24 decided to pause and see if i could figure out an easier way in affinity suite. so first set the view quality to nearest neighbor, then set the view to pixel 1x, then mask a section of the image whose size you know with pixel snapping on (or off if it’s fractionally scaled) with a rectangle shape (not a pixel mask), then you can scale the image either normally or with the point transform tool in designer. once that’s done, turn off the mask, and the entire image is that scale, and then export and select nearest neighbor for resample. unfortunately doesn’t seem like there’s a way to rasterize with a specific mode currently, but they could add that if enough people poke them about it. but either way, you could then rescale based on a specific subset of the image without needing to scroll to see the changes.

Look look. Taran uploaded an hour + long video 🎉

This is EXACTLY what I need holy shit lol

It feels really nice to have a personal gripe vindicated, there used to be a popular emulator that by default would upscale gameboy games using bilinear filtering, and it made every pokemon video on this platform so ugly for like half a decade, the only creators that would notice and fix the settings were like professional editors or film students that did KZfaq on the side

yeah about that nearest neightbor and lanzcos, im convinced that the people who say nearest neighbors is not good and applicable in those instances either didn't read how the algorithm just resize the pixels verbatim and not apply some sampling filter to generate more pixels, or they're not aligning the pixels correctly and get smears from nearest neighbor having to look between the pixels.

Ideally your GRRM screenshot would have also been aspect corrected when upscaling. The image was 640x400, or 8:5, but DOS is generally displayed on a 4:3 screen, meaning the pixels need to be 1.2 times taller. To fix that, you'd want upscale it with Nearest Neighbor 500% horizontally and 600% vertically. (To figure that out, I just typed (640/400)/(4/3) into a calculator and got 1.2. From there you have 12/10 = 6/5.)

Twitter does not support GIF fileformat, so when you upload GIF animation there it gets converted into h264 MP4. Problem is when pixel animator uploads animation where dimensions aren't divisible by 2, the conversion actually downscales it by 1px, which of course will make the pixel animation look even worse than horrible. There's actually way with FFmpeg to force chroma subsampling scaling to be nearest neighbor instead of bilinear so when saving pixel animations as lossy video formats, that's amazing way to technically make them lossless combined with CRF 1/lossless 1 settings. But of course when making h264 MP4 back into images, this helps a lot with reducing colors staying pixelated as well instead of bleeding into areas they aren't supposed to. So what I personally do is I start counting animations pixels, then check how many stored pixels there are. So if there's 10 pixels of artwork horizontally that take 19 stored pixels, that's 190% scale of the original, so [width]/[stored px]*[artwork pixels], I get the original resolution and can just directly scale the image to that without manually stretching it pixel by pixel when it looks OK. Of course sometimes it's hard to count the artwork pixels from improperly scaled or compressed material, but usually it's one or two pixels up or down additionally. After it's properly scaled, GIFs are limited to 256 colors. So if I can find the original palette somewhere, e.g. they posted single frame of the animation, used same palette in past, use console palette, etc. then I can simply use that same palette and properly scaled frames to get back as much of the original as possible. With one animation it had 16 colors, so I literally went and hand picked those 16 colors from areas where they were most likely closest if not original. I did see the four byte burger video after this and felt smart from figuring this out myself earlier already. Looking at the end of selecting those colors, there's clearly limited palette there as well, so I would've also instead of manually selecting those pixels, just try to get the palette that must've been used with that system, index the whole image and force that palette to be used, software should then check for closest match fort the color from the palette and as long as it's not too bad, select correct color automatically for all of those as long as JPG compression was gotten rid of beforehand as much as possible. GIMP allows you to even import your own palettes, so you could just have all most used palettes available and then just image -> Mode -> Indexed -> Use custom palette and select palette. With FFmpeg you would need to use "paletteuse" filter and provide palette as PNG. But the main issue of all of this, is lossy, not idea formats like JPG and h264. I hope we move into more modern formats as whole which make these things less of an issue moving forward while also lowering the size of files, like WebP, AVIF, HEIF, APNG, h265, AV1, etc. etc.

For the redraw, could you use the tools to select the darkest pixels (likely be the lines) then fill them in black and feather/anti-alias it? Could you also use the pen-tool thing

21:46 absolutely love that you're creating feature film-lenght tutorials about making videos look as good as possible, and then either youtube, premiere or whatever gives you a massive fuck you in the form of random data-mosh

BigJPG is a good alternative to waifu2x

"I have a macro for this" of course you do