ぼやけたり顔が崩れる場合はHires. fixもしくはADetailerを使用して下さい.
If the face is blurry or distorted, please use Hires.fix or ADtailer.
*********************************************
好きな衣装や、背景、ポーズを組み合わせて下さい。
Please combine your favorite costumes, backgrounds, and poses.
*********************************************
使用したVAE VAE used
" sdxl_vae.safetensors"
https://huggingface.co/stabilityai/sdxl-vae/tree/main
*********************************************
LORA推奨ウェイト：1
Euler a ステップ30
高解像度補助:R-ESRGAN 4x+ Anime6B,倍率1.5倍、ステップ15、デノイジングストレングス0.5
ADetailerを使用しface_yolov8nで顔修正を行っています。
LORA recommended weight: 1,
Euler a Step 30
High resolution auxiliary: R-ESRGAN 4x+ Anime6B, 1.5x magnification, step 15,Denoising strength 0.5
I am using ADtailer to modify the face with face_yolov8n.
*********************************************
"ADetailer"
(インストールは自己責任でお願いします)(Please install at your own risk.)
顔の詳細を修正する追加機能
Additional features to correct facial details
Description
テスト
FAQ
Comments (7)
How were you able to train this with just a few scenes from one anime episode? Did you use manga art or something?
It's anime only, and XL's learning ability is so strong that you can learn with just a few sheets.
@jibunsagasinotabi That is impressive.
@jibunsagasinotabi Yeah with like 50 Screencaps image Sorted + If Upscale them, you can have something really correct, Tho when more episode will be out, More screencaps/(Or Anime Art that's look Like AnimeTV) = Better Flexibility in terms of Angle/Posing of the Character's and Generally quality too ;)
@KXRO Can I ask you how many repetitions, epochs, and total steps you recommend? I'm learning to get into XL and I still struggle to achieve good results.
@Dazi Hello, Well for me number of Repeats depends of the numbers of Images i did get at the end, Example: with 150+ Images i go generally with 2-3 repeat at Max, More repeats is needed if you got like 30-50 Images.
Epoch generally i'll go always to 10 epoch Max, sometime 8 & Sometime 15 if i want to compare more (Specially for Style).
Steps is deducted from repeat and epoch + batch size, that's why > Generally my target steps is 1000-2000 max at 3,4,5 batch size.
These settings are good for PonyV6 Training atleast right now, work with SDXL Animagine also.
Keep in mind also that one of the most important aspect is the UnetLR (Most Important) and TextLR (More or Less important if it's a Style).
PS: If you are lost on something, HoloStrawberry got a good Starting guide on how to begin (https://civitai.com/models/22530), after this you just need to adjust Settings.
(+ More Advanced Guide for Datasets also if you Search here on Civitai > Wasabiya Datasets guide is one of the best here, It's pratically how i gather/fix a datasets but it's more advanced (https://civitai.com/articles/3996/wasabis-dataset-cleaning-manual) Tho i grab datasets from Various website not only from Booru (EX: ZeroChan, Pixiv, Google, DeviantArt + FanCaps (For AnimeTV), Generally for the Best quality possible). < Very Time consuming, but the best quality at the end Ahah.
@KXRO thanks a lot! This really helpful 🤝
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.