The dataset was surprisingly hard to create; despite having far more scenes than Aika in the anime, almost all the shots of Hakaze were either close-ups of her face or zoomed out so far she is off-model and lowres. I never even uploaded the first version of this LoRA that I created because it was pretty bad.As a result, this LoRA doesn't really have any of her outfits trained in... you can approximate her island garb using sarashi/bandeau/sarong, but that's pretty much it. This version uses the same parameters as the AikaV2.0, but unfortunately shares the same drawback of having a larger filesize, which I believe is the result of the higher network dim. However, when comparing it to a LoRA I trained with the same dataset using lower network dim, this version is pretty clearly superior both in terms of being more faithful and consistent to her face, as well as being more flexible.

Files with SHA256: 10e822ad06980fac62266bd97415f5a4b2704b2f46f19601a1d7675c27bf950a