While Gaia CG works good on some of these cases and some other models also can be usefull when scaling or cleaning up cartoons or old-style anime, this specific usecase has been mixed into other usecase, while needing special treatment…
Frames from footage like this typically has strong black or dark outlines, uniform areas, almost no steady movement of objects (some backgrounds), often repeated frames (5 frames stills, then one frame “moving the mouth”, then 10 frames duplicate frames).-…
Think of old mickey Mouse, japanese 80s Cartoons, etc… Stuff that was mostly drawn by hand and often produced in 12fps or less…
Having a model that aims for exactly these contents would be great. At the moment, in many cases we get ghosting pictures (since the scene change algo misses something) or artefacts from compression or dust or bad transfers are interpreted as picture details and kept. Also, in a case of a lot of repeating frames, all of them are processed, taking a lot of time…
So having some “dedup” detection prior to scaling coul reduce the amount of frames needed top rocess.
A model trained on therse usecases would work like one of the ESRGAN/WAIFU2X Models etc… It probably would be ok to have a single frame aware model here…
But in case it makes sense to look at multiple frames, noise reduction and detail regain could be done…
Proteus - with some settings - can do a pretty good job in some cases - but I beleive a specialized model would provide the chance of even better output and a lot more speed (through deduplication of frames… for inspiration take a look at dedupe filters of avisynth or srestore and similar…)
Before finding this I had just emailed them about the exact same thing. would love a dedicated Cartoon model. I have found some of the existing models are good but depending on the model the outlines can be modified making the characters not look correct. I believe a dedicated model would be able to better edge detect keeping the continuity of the characters between frames.
I’ve made a similar point previously. I’m glad more people struggle with this lol. Maybe they’ll develop a model for it finally.
On the bright side, there are models and software in the open source community that do a decent job. It’s much harder to use, and used AVISynth as a rendering engine iirc, but it works lol. I remember the model I used that worked best was trained on G4 My Little Pony. I used it to upscale Bluey episodes from the DVDs (terrible VHS quality; whoever mastered these should be ashamed) and it somewhat worked!
Have you guys tried Artemis Medium Quality on cartoons? Maybe I’m more forgiving, but I really like all the results I’ve gotten out of that model on cartoons. To give an Idea, I’ve used it on some old Hanna Barbra shows like Scoopy Doo and Smurfs. I think it does a great job of cleaning out the noise and de-pixelizing the lines. The only down-side I have seen is very rarely, it will add detail to backgrounds that probably should not be there. They would fit better, but they’re too small to have been drawn with the same equipment the rest of that cartoon was made with. Either way, not all that distracting, and I only noticed it once for a 3 second scene.
HI. This is the feature request thread, but I´ll answer shortly anyway
The Artemis model was trained on real live footage, so it introduces some artefacts if used on line art old style cartoons. So artefacts are to be expected.
The Gaia CG Model originally was aimed for cartoons, try that one.
Today, Proteus often gives good results when tuned properly.
A specialized Model for old style cartoons is on the wishlist (which this thread is all about)…
I did vote for the model creation. There are so many things it would help with and probably even help improve many of the other models. In the meantime though, I’m just wondering if anyone else has had the same success I have had with Artemis Medium.
I would have never guessed that about Gaia CG from the results it gives on old cartoons. I always thought it was more aimed at modern CG animation ‘cartoons’.
Yes, I have used Artemis models on some cartoons as well - in some cases, the scene change detection missed a few scene changes, so ghosting appeared. But otherwise, good results can come out of this combination.
Gaia CG moved away from the “cartoon only” purpose - and you are correct - it got modern Anime included while training, making it more universal. Sadly, the pure “old style cartoon” purpose was reduced - hence the proposal here
Please add/create an AI model specifically works very well with Anime and Old Style cartoon please. All the current AI models present in the software does not look like it gets the best result for upscaling cartoons and anime. I think it it could be better if there is a decated model just for that
Take a look at HitPaw Video Enhancer’s AI model for animation, see how good the quality is, especially when it comes to lower resolution cartoons or anime (480 resolution) compared to for instance Artemis. That is a model specifically for animation. I would like to have a model like that for Topaz video AI
