ベースモデルはバージョン名参照
約9000枚の画像を解像度2172で追加学習(ver3.1以降)
推奨設定はサンプル画像内蔵のワークフロー参照
base1.0-v3.1
解像度2172で学習やり直し
構図崩れはだいぶ減った・・と思う 多分
変換しただけのINT8ConvRot版はこちら
(botan_animabase1_INT8ConvRot_V3_1.safetensors)
https://huggingface.co/KKTT8823/botan_anima/tree/main/Base1_0
sample画像情報
positive prompt : masterpiece, detailed background, highres, absurdres,
negative prompt : score_1, score_2, score_3, blurry, worst quality, low quality, censored, bar censored, mosaic censoring , jpeg artifacts, signature, watermark, username, normal quality, source cartoon, makeup, simple background,monochrome ,
cfg : 4
Sampling method : Euler_ancestral
scheduler : normal
Description
FAQ
Comments (3)
Were there additional artists added with this finetune? What was the focus?
I focused on high-resolution output and overall stability. I didn’t add any particular artist styles or characters. I prefer to use LoRAs for artist styles and characters, so I tune the base model to stay as neutral and unobtrusive as possible.
@KKTT6783 good decission