>>1023 Is this any better than just Waifu2x?

>>1029 I wouldn't use Waifu2x for anything other than upscaling anime waifus (and even then, SuperYandere might be better). It does fine for MLP fanart, from what I've seen, but it's not nearly as good as the pony model for screencaps. As for photos, it would be silly to use Waifu2x, over say, Topaz Gigapixel. Original: https://u.smutty.horse/mclttowlwnz.png SuperPony: https://u.smutty.horse/mclttpkepbc.png Waifu2x: https://u.smutty.horse/mcmfrmuhyfk.png

>>1033 I usually use Waifu2x-caffe, don't know if it's much different from the website but it does give you a choice of models. I should eventually do a comparison between the models available on caffe and this new thing. Your linked "SuperPony" does look pretty darn good though. Also, can we include DAIN interpolation in this thread?

>>1034 Yeah, SuperPony is pretty awesome; there's a large selection of models in the wiki as well. Lots of room for experimenting with them. >I should eventually do a comparison between the models available on caffe and this new thing. That would be good information since I'm not that familiar with Waifu2x. >Also, can we include DAIN interpolation in this thread? Sure. Anything to enhance the mares that need it. I was planning to go to sleep after I posted this thread, so have another upscale before I go.

>>1035 >I was planning to go to sleep after I posted this thread, so have another upscale before I go. Nice. Also, if you need any bedtime mares before bed checkout >>108.

>>1035 Couldn't get the dependencies to install for Cupscale (probably Windows path length making it more difficult to install torch). But here's some of the models from Waifu2x-caffe. Attached pic is source. >UpRGB https://u.smutty.horse/mcmrknylule.png >UpPhoto https://u.smutty.horse/mcmrkpfcncm.png >RGB https://u.smutty.horse/mcmrkpcztwb.png >Photo https://u.smutty.horse/mcmrkoscpmq.png

>>1063 You know, those aren't that bad, knowing they weren't trained on show screencaps. Also, I need to find out why blemishes start appearing on colors with this model at 4x scale and above. For some images I've been able to keep it from coming up by downsizing an upscale and upscaling the output again, but other times the blemishes pop up no matter what.

>>1070 Good god, it feels like those could have came straight from vector with how clean they are. >Also, I need to find out why blemishes start appearing on colors with this model at 4x scale and above. The bigger the enlargement the more guesses the AI has to make. I wonder if something like jpeg artifacting in the source image could have something to do with it. A tiny blemish at the original resolution might be interpreted to be much bigger when upscaling.

>>1071 I think it might be that, as well as the model just not liking grain or gradients a lot. 1.png was just the original upscaled twice 2.png was the original image resized to 50% and upscaled, then upscaled twice Because the original image has both of those things. https://u.smutty.horse/mcncpdcuukw.png

>>1074 Same thing here.

>>1074 >>1077 Maybe take a very clean png, save it as a jpeg with some amount of compression, then upscale both to see if there's a noticeable difference between the two. Attached is a fairly clean png, a jpeg saved with medium compression, and a heavily compressed jpeg if you want to try it out.

>>1078 I will next. Here's another I just checked.

>>1078 Here's what I got. Could also contact the author of the model and see if he knows something. PNG 1: https://u.smutty.horse/mcndautvdgy.png PNG 2: https://u.smutty.horse/mcndawnfanc.png JPG 1: https://u.smutty.horse/mcndbglggxs.png JPG 2: https://u.smutty.horse/mcndbwgimwm.png JPG (barely) 1: https://u.smutty.horse/mcndcghniwo.png JPG (barely) 2: https://u.smutty.horse/mcndcrohmke.png

>>1080 They're all at 4x scale, by the way.

>>1080 >>1081 Interesting. I don't really see any discoloration of solid colors like we were expecting. I do notice a slight quality drop going from png source to medium compressed jpeg. If you compare the crystal pony in the background of both you can see her become a little less defined in the jpeg. But over all they are very similar. RD seems to be mostly unaffected. (barely) jpeg is as expected terrible, don't know if there's many conclusions to be drawn from that. Also, what's the difference between 1 and 2? 2 seems to be consistently worse. Lyra and Berry at the market stand seem to lose their lines in 2.

>>1082 >I don't really see any discoloration of solid colors like we were expecting. The discoloration is there, Anon. In the PNG image, it's on Berry's tent roof, Rainbow's closest hoof, and the blue section of her rainbow mane (not the outline), and the orange part of her hair, to name the ones I spotted. >Also, what's the difference between 1 and 2? 1 was just the original upscaled twice 2 was the original image resized to 50% and upscaled, then upscaled twice 2 is usually worse, yeah, but flipping between 1 and 2 can help spot the discolorations.

>>1083 I think I see what you mean now. The entire color seems to be a slightly different shade. But I don't see as much of the splotchy discoloration seen on other pics. Looking very close at the left hoof I can start to see some, but on my they are rather difficult to discern.

>>1085 I think the splotchy discoloration occurs every single time, but it's more noticeable in certain colors. For example, it's especially easy to notice it happen with Twilight, but it's extremely hard to get the same results with Fluttershy. Pic related was upscaled to hell and back.

>Realized I haven't posted any DAIN content yet A lot more finickily than I remember, but here's a Twi eating some burgers.

>>1086 I definitely see it here in the hair. The splotchyness is a lot more noticeable than the slight shifts in color. Fortunately it seems like most of the time it's not too bad.

>>1087 Wow, that looks very smooth. >>1088 I say that, because with a similar image and scale, the splotchiness is a lot more noticeable on Twilight. https://u.smutty.horse/mcnedcjubto.jpg

>>1087 Doesn't look as good as the original.

>>1097 I was really surprised at how DAIN was handling things. I remember it being a lot more impressive. Most of the stuff I tried now just comes out real blurry. Since I still have a copy of v0.2 maybe I'll try and compare it to v1.0 and see if things really have gotten worse.

Thanks for this tool anon. It will be useful in upscaling some textures I am working on.

Here's a challenge: I'd like to get this nice pony themed tarot deck to a functional resolution. this is the largest I can find it in, but unfortunately it's a bit small to be usable, and my only computer is a potato laptop so I can't upscale things myself. https://u.smutty.horse/mcpalfzirid.png

>>1439 What do you mean by "functional resolution?" I'd love to give it a shot, though. I'll try it out later tonight; I'll check if there's anything to be gained by cropping all of the tarot cards by themselves and upscaling that instead as well.

>>1023 Link to the archived thread: https://desuarchive.org/mlp/thread/37338910/#37338910

>>1439 I'm sorry I took so long, Anon. I got really busy the past few days. Here you go: 2x: https://u.smutty.horse/mcqygwqrihg.7z 3x: https://mega.nz/file/uIAQnJAI#P2v0ukM0v-OWoeEONxECE_6S-IyY_ZDbHym71Fkszf4 I couldn't do 4x with Cupscale because it has an image size limit of 8000 x 8000. if the above links aren't good enough, I can try cutting out the cards individually and upscaling at 4x then. Let me know if these aren't good enough.

New SuperPony models are out since I last checked! Loyaldk-Super/Medium/LitePony_500000_V2.0: "Upscale MLP episodes. Full version. Able to handle compression better compared to V1.0 and no longer creates halo's/rainbows. Good for Vector 2D art however converts detail to blobs." Here's a quick comparison. All at 2x scale. 50% resize with 2x -> 2x -> 2x I don't like the new version by itself at all, and I think the old version is superior as long as you fix the coloration in Photoshop. Using them both together might have some merit, though. Mind you this was just a quick little test, so more can be gotten out of it I'm sure. Original: https://u.smutty.horse/mcyfrjnkuix.png Old version (V1.0): https://u.smutty.horse/mcyfrxeisqn.png New version (V2.0): https://u.smutty.horse/mcyfrzsruou.png Old combined with new: https://u.smutty.horse/mcyfscdqmdm.png

>>2533 The blemish is still there with the combined version. But, at least you can upscale it further, without even more blemishes appearing.

>>2533 A definite improvement I'd say. Eventually I got to actually get this working on my computer.

I feel like a bit of a brainlet, I'm just content using Waifu2x caffe for now. This looks complicated.

>>2664 The hardest part is installing it, Anon. After that, you can upscale to your heart's content. Besides, you can always request for images to be upscaled ITT if you're ever not pleased enough with any Waifu2x upscale or just want to see how well some of these other models do.

I wanted to revisit one of the images that was upscaled in the original image upscaling thread on /mlp/. Pic related is the original image. Also, for whatever reason, Smutty is making the images significantly darker for some reason. I had to use Imgur. 2x Here's what OP finished off with (realsr): https://i.imgur.com/pgrmtAy.jpeg 4x 50% Realsr -> Deviance (new): https://i.imgur.com/P7wn73E.jpeg 4x 50% Gigapixel Standard (no noise suppresion) -> Deviance (new): https://i.imgur.com/geRR6nm.jpeg

I'm also starting to really warm up to the 2.0 version of SuperPony. These are the original, V2.0 4x scale, and V2.0 4x scale -> %50 V1.0 4x scale respectively.

>>2741 >>2742 These all look pretty good.

>>2743 They're amazing. Being able to upscale almost anything is exceptionally wonderful. I'm completely enamored by this one as well: https://u.smutty.horse/mdbfarodnqh.png https://u.smutty.horse/mczxdeariys.png

>>2741 Nice

Just another quick comparison, I haven't had the time to play around with the SuperPony models yet. I've also never touched the liter Pony models, but I don't really see the point. It goes original, old model only, then a combination of old and new, all at 4x scale. I forgot exactly how I got the second image (I did that a while ago), but for the third one I just went new model at 4x -> 50% resize with old model.

>>2944 It's looking pretty clean. I wonder if there are any old pony games that could be remastered with this?

>>2945 Do you mean like upscaling all of the assets? That would be pretty neat to see, although I'm not sure what you could do with the pixel art ones. You can't really add more detail unless you do it manually, and IIRC all of the current pixel art models only smoothen out the image or stylize it like a digital painting or other. I've thought of taking an MLP episode and upscaling every frame with the SuperPony models, but in such a way that it looks trippy or weird. Probably compress the frames to hell and back to around 50 x 50 and then upscale the shit out of those images. They tend to resemble what the show looks like, only super weird and with blue highlights.

I wanted to post some more great upscales. Links are pics-related, respectively. All are at 4x scale. Gigapixel Standard -> Fatality Comix https://u.smutty.horse/mdgzdwfvwbw.png Gigapixel Standard: https://u.smutty.horse/mdgzdeccbei.png Fatality Comix: https://u.smutty.horse/mdgzecbsexp.png

>>3180 Again, pretty good with minimal artifacting.

I've also been meaning to post some of these upscales as well. Gigapixel's noise reduction is pretty amazing. I've yet to try some of the other models for doing that, though. With that said, Gigapixel is pretty great for adding some detail so you can resize it down a bit and upscale from there, and these images needed the help. They're still not perfect, but using some of the models by themselves gives less than desirable results. https://u.smutty.horse/mdgzdmjcbzz.png https://u.smutty.horse/mdgzbyxnuql.png https://u.smutty.horse/mdgyzdrprhy.png

>>3195 Although it could be argued that the noise is part of the artist's intention, like in >>2741 but I don't think that was the case for the first and third images.

>>2741 I'm reuploading one of the 4x scale images to Smutty. The darkening may be entirely on my end, because whenever I download the images, they look how they're supposed to. That, and I learned Imgur compresses images. https://u.smutty.horse/mdbehpfnopy.png

I found a clear example of the V2 Superpony model being leaps and bounds above V1. Pic-related is the original. These are at 4x scale: SuperPony (old only): https://u.smutty.horse/mdjgjvwseeb.png SuperPony (new+old): https://u.smutty.horse/mdjgjddrbet.png SuperPony (new only): https://u.smutty.horse/mdjghfggrdq.png

>>3590 Cool stuff.

I wanted to post an interesting little upscale I've done today. Pic-related is very pixelated, and the SuperPony models couldn't handle it very well. Until I used "4x_scalenx_90k." It smoothed everything out quite well. https://u.smutty.horse/meemwvdblba.png Scalenx -> SuperPony (new): https://u.smutty.horse/meemwwgaykp.png

>>5320 Very nice. This post both reminded me that upscaling is a thing, and that newer, better models more suited for pony have been made. I think the last time I had upscaled anything was like a year and a half ago. I updated to the latest version of cupscale, grabbed some new models, and gave it a quick test with what I think are great results. Take this Chrysalis pic for example. It was made by one of ((those)) artists who don't release their full artworks, instead hiding them behind multiple layers of annoyance. In this case, the annoyance in question is DeviantArt. The thing about DeviantArt, and what this artist does, is that you can set it so that people cannot view or download the original file. Instead, the site serves you a copy of the original that has been scaled down and had the quality reduced by 25%, introducing artifacts and pixelization. Back in the old days you used to be able to browse the page source to bypass this, allowing you to grab the original file, but apparently this either doesn't work anymore or I'm too stupid to figure it out. What I did manage to figure out, however, was to get the downscaled image, but at 100% quality. So I took that, scaled it up with SuperPony 4x V2, then manually scaled the image back down to the artists original dimensions of 3508x2480. While I was at it, I gave the image a little edit. You can see in the original that Chrysalis has a bit of a frown at the corner of her mouth. I don't like that, it makes her look unhappy with the situation. A slight adjustment gave her a more neutral expression, hinting at a smile. Changes the whole feeling of the image in my opinion.

>>5345 Upscale came out looking pretty good. Your edit, while subtle, I think is very effective. Nicely done.

>>5320 Looks great! >>5345 Oh wow, that's really nice!

>>5345 It's pretty amazing how well the SuperPony models do with fan art. We have more tools to deal with bad decisions from artists or just low-resolution fan art in general.