>>8737690
well, after dicking around with it a little bit, about all I can really say about it is that it uh, sort of works. Sort of.
https://www.mediafire.com/folder/k58cm0mq3nspr/aoi_tiduru
same deal with the others with jz235 as an activation token.
it's got a lot of issues and it's mostly due to the dataset.
Issue #1 is that it isn't even very stylistically consistent. Even in the nimi image set it swaps a bit between a more sketchy style and a solid linework style. Also that image set has a different shading profile than the two images with a white background(which are super important for reasons).
Issue #2 is that that art style doesn't really like to be downscaled and you lose a bit of the linework in trying to do so, so it's better to work with crops and do minimal downscaling if possible.
issue #3 is the obvious of the dataset being really small.
issue #4 is honestly that it's fairly generic-ish digital art and that makes it really hard to get a style to express since it wants to just get lost in the model.
anyway, as it is it has issues of wanting to do a lot of closeup shots due to the uh, method I chose to append/fluff out the dataset(which you can find here https://files.catbox.moe/ufvs5f.zip ). It also has an issue of hardstuck backgrounds, also for the same reason. Though you can prompt white background and it'll at least listen and give you an empty space, even if it's sometimes off-white instead of white.
You can also try and run t2i at .6~.8 weight to try and get it over the close-up syndrome and then use full weight on upscaling. Or just do t2i in a similar but different style and i2i upscale with it enabled, idk.
>>8737818
it's using l2 loss. the huber values are there as a holdover from some random testing(the config this is based on is like a year and 8 months old) but shouldn't actually be doing anything since it isn't set to huber or smooth l1.