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-fp16-fix”
" sdxl_vae"
https://huggingface.co/stabilityai/sdxl-vae/tree/main
*********************************************
LORA推奨ウェイト：0.8～1.0
Euler a ステップ20～30
CFGスケール 5.5～7
高解像度補助:R-ESRGAN 4x+ Anime6B,倍率1.5倍、ステップ5～15、デノイジングストレングス0.5～0.6
ADetailerを使用しface_yolov8nで顔修正を行っています。
LORA recommended weight: 0.8～1.0,
Euler a Step 20～30
CFG Scale 5.5～7
High resolution auxiliary: R-ESRGAN 4x+ Anime6B, 1.5x magnification, step 5～15,Denoising strength 0.5～0.6
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 (18)
Deja vu! I've just been in this place before! Higher on the street and I know it's my time to gooooo~!
Nice model! And thanks, you jerk. Now that song is going to be stuck in my head for the rest of the night! Haha
Why is test in parentheses? Pursue the most perfect version?
Learning about cars was difficult, so I took multiple tests.
@jibunsagasinotabi Pony is prone to misjudgment at long distances, strengthen training on people and cars at close range. If it is too far away, pony may be difficult to understand the the anime character relationship with the car. Superfluous background will miss the point. alpha and dim can be trying 3:4 or 1:1.
@ghostpaint
I've been trying a lot of things, but it's not working
@jibunsagasinotabi It is difficult to train people and car. This situation happened to me when I was practicing lora before. I just relied on cropping or eliminating unnecessary interference. If the alpha is too low, it is also difficult to train multiple people and objects. Of course, your everything has been done well. Or train two lora separately, use them simultaneously.
@ghostpaint
Increasing the network alpha increases capacity, so we try to design it to be as small as possible. If possible, we would like to complete it with one LORA.
@jibunsagasinotabi Alpha does not increase capacity, dim does. high alpha is suitable for establishing new concepts. If 1/2 is a fastball, 1/1 can be imagined as a breaking ball. For example, if I want to train a character, simple daily postures, 1/2 is enough. But if I want to train his special move at the same time, I will try to increase the alpha.
@ghostpaint
I understand that I can change the DIM rate and set it to the same number as Alpha or a number greater than Alpha, but the learning value will decrease, so I need to increase the number of steps. I'm using the current value so that it can be completed in 300 to 500 steps if possible.
@jibunsagasinotabi It is possible, high alpha is not recommended to only train the character, and the stability is not good. High alpha also has high requirements on material quality. It may be too late to correct it and the training is over.
@ghostpaint
Yes, when I was using SD1.5, I was creating it like that, with high Dim, low learning value, and high step settings, but the PC specifications were not enough for use with XL, and it was not realistic to take several hours to create one LORA.
@jibunsagasinotabi I am vram 16G, use adafactor to train pony lora. With some settings to reduce VRAM requirements. If I use adamW, 512x768 is the limit. But it can use 1024 x 1024 in adafactor. In addition, the best setting for SD1.5 is 768x768. I once used 1024x1024, but the results were not better. Foreign experts have also tested it and found that 768x768 is the best setting for SD1.5.
@jibunsagasinotabi The adamW has a low learning rate, low batch size, high DIm, and high number of steps. Right now, except for celebrities, there are fewer and fewer people training like this. Training too long...
@jibunsagasinotabi The shorter the training time, the more times you can adjust the material to increase, improve more versions. On the contrary, after AdamW was trained, many people still reluctantly released it, quality varies, and it was difficult for most people to do it again.
@ghostpaint
When I was using SD1.5, I also studied with adafactor and 768x768. With XL, I use Prodigy to maximize the learning efficiency.
@jibunsagasinotabi It should be the most advanced optimizer currently.
bravo
素晴らしい作品です。
thank you
ありがとうございます
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.