Anima-H3-Booster-Lora

This was made raking the Anima Wai full model across the text encoder layers of H3 in blocks 0,1. Each anima block was cycled over them 28 times each to a fairly extreme linear magnitude and shift. That was extracted at rank 2048 (100%) to make this. Main block grafting with Anima produced worse results than other models I've tried with so it was left out. I'm not exactly sure what worked, why it does what it does. I don't know if much new data was transferred to the model, or if this lora just fully alters tokenization and the model behaves differently, but in t2v the results are substantially effected by style, and i2v there are some big detailing improvements. H3 is actually very anime capable, you can take booru or plain language anime image prompts and use them directly as the compositional instruct for the first shot, or generated anime live wallpaper outputs fairly easily. Prompting is quite important to reach new effects, random booru tags can be used inside prompt as well.