Made during experiments with SuperMerger. Works either on clip 1 or 2, from 1:1 to 16:9 ratio. For first version use realistic tags to get more modern anime looks instead of more classic flat one. Latter ones have more distinct features that mostly comes from further and different experiments.
Prompts wildcarded, so some can bear little sense.
Description
FAQ
Comments (6)
Looks great. Which models are merged here?
More then ten really, but with specific methods, or it be just latent mess.
Could not remember what models exactly but there are three main groups - modern anime, flat anime, cartoon-ish anime, into each of model i injected data ("train difference" on 0.5 weight) three specialized ones (though only pair survived all three injections, mostly on further step taken stage 1 and 2 models), then selected ones (not all) merged inside group, and then grouped models merged to create resulted, as seen in metadata of second image of first post.
TrainDifference is variant of AddDifference from SuperMerger extension what allow to inject one model into another and get results similar to finetune source model with difference data.
@eliont For a 10-way model merge this looks really great. Thanks for the detailed explanation. I was basically curious to learn which flat models were part of the merge since I'm always looking for a nice, oldschool 2D-flat aesthetic and your merge definitely retains this vibe.
@ratherbeanoutcast recovered some info from logs - YiffyMix, momoko (not momokoS), ExpMix_Line, CuteYukiMix (idr what version), ChilldrenStories (anime), Cardology Mix, it's not all, ofc
@eliont Thanks. I played a bit with your mix, definitely clear-cut distinction when using realistic vs. anime - that's really nice. Couple of suggestions for the description so that people are aware:
1. Seems to generate a lot of sketches/manga 4koma-like panels. Throwing monochrome and sketch in negative prompts helps.
2. I think you should mention the suggested VAE. I assumed it was baked-in, but it doesn't seem to be the case.
BTW. The model responds really well to even minuscule changes in prompts. It's refreshing to see.
@ratherbeanoutcast Thanks.
1) I always have "monochrome" in negatives. Sketch can sometime give interesting results.
2) No, VAE not baked in. I used ClearVAE for this gens, bit you can take any you like.