Trained on 315 images of Island Fox (no repeats) from Kemono Friends on Anything v3.0 Pruned FP32 (however it works on every model)
This time I added the tag "shimahai" to the training so that it associates "shimahai" with many features of her, and fixes the association with "fox" and "kemono (in parenthesis!)" messing up the generation
Of course, island fox (kemono friends) still works
Vtuber art is filtered from the dataset, that will be its own model
Stats:
315 512x images (no repeats) for 20 epochs
Network dimension: 32
Base model: Anything-V3.0-pruned-fp32.ckpt
Batch size of 3 on RTX 3060
Generation tags: shimahai, 1girl, ((solo)), animal ears, virtual youtuber, short hair, fox girl, fox ears, fox tail, gray gloves, gray pleated skirt, white necktie, twintails, black undershirt, grey hair, orange hair, red blazer, white loafers, red beret, yellow eyes, microphone, long sleeves, fur trim, black belt, bangs, hair between eyes, orange legwear, two-tone hair, black hair ribbon, frilled skirt, sidelocks, breast pocket
Sample images are 512x512 latent upscaled to 1024x1024 and generated on Anything v3.0 Pruned FP32