사용자 LoRA
SD
_dabdabil
This LoRA is based on the works of _dabdabil (aka Dabi_Bill). It was originally trained by Noise_Connoisseur on Civitai.
Dataset: Manually sorted and cropped 352 images, 896x896 resolution. Trained on low speed with 1 repeat for 18 epochs on "AnyLoRA" checkpoint.
Testing: Overall testing took upwards of 6+ hours. I took great care to test for overtraining and settled on epoch 16. Tested both anime 2D and 2.5D checkpoints. Tested with other LoRAs, embeddings, etc.
Strengths: Aimed at generating teens/young adults. Facial expressions are both diverse and easy to prompt (very close to original ones). Exceptional at generating portraits, bodyshots. Easily handles both SFW and NSFW. Compatible with almost any other LoRA, negative embedding, etc. Can be used to remove the "porcelain doll face" found in some 2.5D models (e.g. REV Animated).
Weak points: The style itself is very sketch-ish/simplistic. This means that it will simplify the output. If you are using a very low detail Anime model (Grapefruit, Orange Mix, etc.), you may (in certain situations) experience lack of detail. This can be aided by either lowering the strength from 1 to 0.8, using BREAK command in the middle of the prompt, or by using the "Add detail" LoRA. When used with higher resolutions it may occasionally (5-10% of the time) generate speech bubbles or comic strips. This is caused by the dataset and may be improved in subsequent versions.