https://nyaa.si/view/1801041https://nyaa.si/view/1808448Second batch with 26-50 is out now. Full details about the upscale process are in the Nyaa posts, but basically, this is an upscale of the original SD version of SEED. It was done using a custom upscaling model trained specifically on the HD Remaster.This version is based on the SD Blu-rays from Rightstuf, so the source is much cleaner than the original DVDs, and it's from hchcsen's remuxes, so they're as good as you can get.
>>22595067SD version upscaled 200% for comparison
>>22595067Would this work with Eureka Seven? Train the AI upscaling model using the later Eureka Seven releases to upscale the original show to 4K then downscale it back to 1080p?
>>22595077Would you believe I've actually been trying this? The model "works" on E7 but the improvement is negligible imo. There are other models that definitely improve the sharpness dramatically, but they enhance and smooth the background details too unnaturally to be useable.Needless to say, training models has been a rabbit hole. I've been looking into good training references for the Dynit Sailor Moon DVDs, since my initial version with Topaz was lackluster, looking back on it.
>>22595079Is it effective on cel shows? I know there's still some old mecha that are still in SD like L-Gaim. What would happen if you upscale it with models trained from the Bluray rips of FSS or Zeta Gundam (since it's they're aesthetically similar to L-Gaim)?
>>22595095The thing is, the model has to be trained on the low res and high res versions of identical content. The problem is finding HD versions that are "good" comparisons. Unfortunately, there are many HD remasters that are overly denoised, so the model then becomes trained to remove a ton of background detail.I've tried creating synthetic downscales of 16mm content (G-Gundam), effectively making my own pretend DVD encodes, then training against the original BD version. I then also added some 0080 DVD and BD comparisons into the training to give it some real world variety, and it's pretty effective, though it can smooth some details too. The 0080 training actually goes a long way towards improving line art though. Let me grab some test screencaps...
>>22595100Okay, so here's a raw DVD screencap. I have a few models I've been working on with varying results.
>>22595111(In case it wasn't obvious, that was upscaled to 1080p for comparison)Here is my current "hybrid" model, which combines training data from the synthetic G-Gundam downscale, and 0080's DVD and BD versions. It is definitely smooth and perhaps overly sharpened to the eye, but it is the most consistent so far.
>>22595112This is with the synthetic G-Gundam training alone, which retains the grain, and is naturally somewhat softer, BUT it also sharpens the lines in very inconsistent ways at times. It also doesn't alleviate issues with the older transfer like the chromatic abberation and color bleed. On the other hand, the 0080 training data helps a lot with that, since it compared an older transfer to a newer one.
>>22595114Lastly, here was my idealistic attempt. Knowing this was a Toei production from around the same time, I figured there is no better comparison to use than the DBZ Dragon Box and the abandoned DBZ Level Sets, which were full of grain.The problem was that I was unable to find a BDMV or Remux of the Level Sets, so I had to train against a slightly denoised encode. This unfortunately tainted the training, so it removes a shit ton of detail from certain background shots, but it otherwise has a really pleasing look to the outlines (though it doesn't do much anything for the color bleed).
>>22595077If you're curious, here's the raw BDMV of E7 (US release. I'm not sure if the JP version is any better).The upscaling is aliased and edge enhanced, so it's difficult to ascertain the exact source resolution.
>>22596044And here's with the SEED upscaling model applied. I've also removed the edge enhancement and done a mild denoise before upscaling since the BD encoding isn't great.
>>22596047Well, well, well. How's this?
>>22597344This looks amazing. How does it look when it's moving? Any quick WEBMs?
>>22597363https://mega.nz/file/7F5hnQqR#t1TNVfvNkrAe_BcJol2J0u4mlIdDtSFjZH4sBNBtkxYQuick encode of the OP. The credits definitely get fucked from the upscaling, but I think it would be forgivable in the grand scheme of things.
>>22597380This is how the Blurays should've been
>>22597344Holy shit this is insane, and I thought the show couldn't look any better.would definitely pay for Blu Rays in this quality.
>>22598039>>22598077Yeah I'm honestly still blown away by how effective it can be. I've been on a fucking spree training models the last three weeks for both digital and cel purposes.
>>22598263Is it me or is it taking out a lot of the subtle shading and making it look too pixel perfect painted now? Like you oversaturated an image with a glaussian blur or something.
>>22600707I mean, it's never going to look perfect. There are quite a few models out there that make it look REALLY pixel perfect - like promo art level. Most of those tend to fuck up the blur effects used in backgrounds too.
>>22600946Oh, no, I didn't expect it to be. I'm just curious if there was anything super wonky with the upscaling (mainly with background stuff)
a heads up if people are still looking at this threadthe ever excellent ai upscale model for older web and DVD release had finally releasedhttps://github.com/Sirosky/Upscale-Hub/releases/tag/AniSD2x_AniSD_G6i1_SPAN_215K and 2x_AniSD_AC_G6i2b_SPAN_190K is what author recommend for most case,and with JaNai GUI running with tensorRT,it is ver possible to upscale video in ver high fps.
>>22602037I've been following Sirosky's guides for this whole endeavor, and these models look legit. I can't believe I'm saying that they almost seem *too* faithful just judging by the screenshots, but that's awesome that he achieved it. I'll have to try these out sometime today.I haven't played with JaNai much (I should, it would open up more models to me), as I'm always loading these in via Vapoursynth for other processing, but this might give me a reason.
>>22602298Here's the same E7 frame with AniSD. It's nice, but it doesn't feel like enough of a change to justify the compute power.
>>22602346Imo, the ideal is somewhere else. The quick model I showed in >>22597344 has artifacts in a number of places, but there's gotta be a middle ground....
>>22595067I've watched your CCA and Big O upscales and have them in my personal digital library; if you don't mind me asking, how the heck do you make these? I'm a programmer and I've been dabbling in AI image generation so I'm not a complete layman, but I'm interested if you have any resources you could share on how to get started doing this kind of stuff.
>>22606992EG is the OGHe's been doing it for 20 years.
>>22606992He did a blog back in 2014 if that helpshttps://rxmastering.wordpress.com/
>>22595067RX, question: how does your upscaling model do with text? Even my dicking around on Topaz doesn't know what to do with it.
>>22607283Same shot from the official HD version.
Well... You have already started E7 sooo......
