Description
Initial.
The LoRA was trained on 96 frames extracted from the ending promotional video (ED PV).
Captions were generated using SmilingWolf/wd-eva02-large-tagger-v3 with a threshold of 0.35. Incorrect tags such as colored_skin were removed, and character names, camera angles, and other essential details were minimally supplemented.
No natural-language captioning was performed. However, natural-language descriptions in prompts may still evoke compositions from the training images.
LoKr, dim32/alpha32, CAME, UNet LR 3e-05 cosine 1,000steps, 96 images (512² res).
FAQ
Comments (1)
Awesome style!