Miranda Wright from Bonkers (1990s). I wanted to capture the Disney afternoon animation look - so I upscaled the references and took specific shots from the episodes. I think this turned out very well. There are some specific models that will retain the western/disney look over applying anime, but overall very compatible. You can also apply LORAs with very little style loss many times.
Description
Comments (5)
Having done training off a lot of 480p cartoons/video game cutscenes, I recommend not upscaling images if you can meet the minimum size requirements (256x256), and if you absolutely have to to meet it, use next nearest. LoRA making teaches patterns and when its presented with a bunch of upscaled images, it learns the upscaling artifacts are "supposed" to be there. Also the training "sees" images as a depth map, which is comparable to (but not) a vector image, so even in the best case upscaling training data is redundant.
I hear you - the only reason I did that, was because I had crappy results trying with originaly resolution. Even having the episodes on hand, one forgets how 480p television used to look (like crap). It just didn't do well.
@superkat What kind of "crappy results" were you getting?
@NanashiAnon Face was off a lot. It really was just with this model, hence I did upscale and sharpen the images used.
@superkat Ah, then it most likely learned that "full body" shots in the work have low face detail and reproduced that. Face fix can fix that even in extreme cases (as an experiment, this model was trained on 34 images, only 9 of which had a real face and the rest had ~15x15 blobs at 4x size or no face at all, but face fix still works). If you want to avoid needing face fix, I'd recommend mixing in close ups if possible, and only upscaling the distant ones that actually need it and tagging those "upscaled".
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.