v0.0.0 (Experimental / Test Release) This is a test LoRA created using Civitai's online training platform with just a single training image. The image caption consists entirely of natural language descriptions.
According to my test using Civitai's default settings:
At around 600 steps, the model began to memorize the training image.
At around 1000 steps, the model fully converged.
Description
Initial.
dim32/alpha32, AdamW8bit, UNet LR 1e-04, constant 2000, 512² resolution, 1 images.
Comments (1)
かわいそうなのは抜ける。
ただしキャラによる。