What does it need to train a non-compact model to a "into" a compact-model?

[Platos]

What would it need to train a non-compact model "into" a compact model (or lite, ultra lite etc.) ? So i mean basically make from a non-compact model a compact model.

What would you have to do ?

[TreeTops]

I don't buy Bud lite. I just couldn't resist that pun. But, I have no idea what you are asking.

[Platos]

And i have no Idea what you are talking ^^

[TreeTops]

It was a U.S.A. joke.

[Platos]

Ah, i don't know that

[Selur]

https://www.youtube.com/watch?v=l_tZE5l90VY might help,..

[Platos]

Thanks, i saw this only today. The video seems to be very good. (He is swiss probably )

Im watching atm, but what do you do, if you don't have a high quality image of your desired Anime ? So maybe this is just an old Anime where there is absolutely no "high quality" image ?

Do you have an idea? Probably im forced to use another Anime with similair artstyle but higher quality, right ?

Or do you have a better idea?

[Selur]

The main point of these models is that they do not have to be trained with every available anime/whatever, but that they are trained with some reference data and some degraded data and then create rules on how to create the good data from the reference data. The model only works on portions of an image and thus rules created from the training set ideally allow handling any similar content.
Those models from https://upscale.wiki/wiki/Model_Database often use as good data sets from https://upscale.wiki/wiki/Dataset_Database and create the degraded version themselves. Alternatively, folks create a new good data set (from videos&co) and then degrade it.
I never had the time and motivation to figure out how to train models.
The upscale.wiki folks also have a discord channel: https://discord.gg/cpAUpDK with new models, folks over there probably can give you better advice on how to train models&co.

From what I know, you can't convert an existing model into a compact model, but you need create a new compact model.

Cu Selur

[Platos]

AH ok, yes thank you.

Yeah but it seems it does not take soooo much time (so "render" time). In the Video he said sth from 24hours and he said also he has a 8GB graphics card. Therefore it can't be a better GPU than mine because i know the models. So im fine with 1-3 Days.

He did btw. share his trained model here: https://github.com/Phhofm/models

I tried the LD-AnimeJaNai. It does a better job in "not washing out" textures, but it makes white lines around the black lines (which does LD-Anime too, but not so much).

I tried then also 2xHFA2kCompact and i like this so far. But i have to make further tests. Maybe someday i will do an own compact model if i dont find a good one (LD-Anime does not work great on a source im currently upscaling).

[Selur]

He did btw. share his trained model here: https://github.com/Phhofm/models

Interesting.