Multi Lora 20 styles in 1 SDXL and 1.5 for animatediff
check out my video, i used some of the styles, it was before i trained the multi model and now its so much easier to do stuff like this )
https://civarchive.com/posts/6346304
It started as an experiment in sd1.5 for animatediff, if instead of changing loras while running animation, it would be possible to train one model on many styles, the results turned out much better than I expected and I came up with 20 styles in 1 model and i decided to train it also for sdxl.
my conclusion regarding vid2vid is that a good approach can be to train one model in advance on all the styles we want In our video in each project.
its better to improve each prompt according to the style you want, but this is a basic prompt to start that can work on all styles:
positive: #subj in a human shape dancing in the street, made only of #subj, cool and crazy photograph of a #subj in the form of a human, high detail photo, ultra sharp focus, cinematic, artistic #subj concept art, balanced colors, high detail photo, ultra sharp focus, cinematic, artistic #subj concept art, beautiful urban background, balanced colors, high quality photography, 3 point lighting, flash with softbox, 4k, Canon EOS R3, hdr, smooth, sharp focus, high resolution, award winning photo, 80mm, f2.8, bokeh
just replace #subj in any one of this: balloon, soap bubble, cactus, cotton candy, construction crane, cream, broken egg-splash, electric street light pole, fire, gummy bear, tree made of tree trunk, ice, origami paper, paint drip, Spaghetti, yellow toy bear robot, snail, liquid water splash, skeleton x-ray, garden flower made of leaves and a flower as head
Negative: (remove the illustration, sketch, painting, cartoon... if you want illustration style)
text, signature, blurry, low quality, out of focus, out of frame, low resolution, pixelated, ugly, illustration, sketch, painting, cartoon, dark, bad light, over saturated, burnt, digital art, human being, skin, hair, face, clothes, outfit, nude, naked, nsfw
you can also grab the prompt from any of the example images
*for more affect and more control on the output you can combine the lora with ip adapter loaded with a good reference image of the style you want