Note:
To achieve better results, it is recommended to use lora block weights.
A simple method to set the weights is as follows: CHARACTER:1,1,1,1,1,0,0,0,1,1,1,1,1,1,1,0,0
BACKGROUND:1,1,1,1,1,1,0.2,1,0.2,0,0,0.8,1,1,1,0,0
triggle words :
Hoshino Ai, long hair, purple hair, streaked hair ,purple eyes, star-shaped pupils, hair ornament,
model:
most of anime-like models
Version
V8.3 includes improvements from Version 5, offering slightly enhanced brightness; however, its performance remains subpar when it comes to hair ornament details.
V9 employs the V1 dataset and has been re-tagged to eliminate the yellow cloth around the neck, resulting in a more satisfactory output.
V1.1 has issues with red pollution, frequently displays yellow cloth on the neck, and its posture is relatively static.
V5 addresses the issues found in V1.1, but it introduces some new problems, making it not necessarily superior to V1.1. After making significant changes to the dataset, it appears that there are too many images with poor lighting, which may occasionally compromise the image quality.
V5 appears very dark, and it seems challenging to increase its brightness, and if you use easynegative, it will somehow look like 'deep-frying' without highres fix. I'm trying to solve the problem.
So choose the model you like.
It seems civitai can't read the pnginfo from my picture due to Dynamic Prompts.
Description
dataset: pixiv until 4.19 +anime 1
FAQ
Comments (3)
"V5 appears very dark, and it seems challenging to increase its brightness, and if you use easynegative, it will somehow look like 'deep-frying' without highres fix. I'm trying to solve the problem."
Might trying lowering the Unet learning rate a bit.
Trying using the auto1111 extension "sd-webui-additional-networks" to create an X/Y/Z plot of different TEnc and UEnc weight multipliers, so you can see what ratio of Text/UNet gets you the best result. :)