For v0.0.0, I trained the LoRA using 24 training images featuring the same pose and the same room, plus 6 additional images showing variations of the pose and environment.
As a result, applying the LoRA tends to overfit the pose and scenery, causing both to become strongly fixed.
The training used the terms "virtual_insanity_pose" and "liminal space" in advance, but during generation, the model tends to reproduce the training images almost verbatim regardless of which trigger words are used.
2:40～
Description
Initial.
dim8/alpha8, UNet LR 1e-04, AdamW8bit, warmup 0, constant 600 + cosine 400 steps(100/300steps), 512² reso, 26 images.
FAQ
Comments (1)
最近このyoutube動画見た。なつかしいですね。当時カップヌードルのTVのCM見てすぐ輸入盤を買いました。なんとなくしか歌詞の意味がわからなかったんですが、その動画で自分の英語解釈はだいたいあってたと…(笑)。
Loraは頂いていきます😁