Kohaku XL alpha
An anime SDXL model trained on 100kπ images.
Introduction
This model is trained from sdxl-1.0-base and merged with some other anime models to get optimal result.
I'm planing more training on more images (around 1.6M imgs) and some style lycoris models. so now it is just alpha version.
Usage Details
This model is very flexible on resolution, you can use the resolution you used in sd1.x/2.x to get normal result(like 512x768), you can also use the resolution that is more native for sdxl(like 896*1280) or even bigger (1024x1536 also ok for t2i).
recommended negative prompt for anime style:photorealistic, 3d model, cinematic, bad anatomy, blurry, disembodied limb
Acknowledgements
Datasets
280k imgs from Mynefactory
30k imgs from cybermeow
Models
Description
nya!
FAQ
Comments (20)
What Vae do you use with this? or is it baked?
The sdxl fp16 fixed vae
1.6M images? I think that would be the closest to NovelAIs 5.3M images interestingly, of course if this Twitter post is to be believed at least. Be really interesting to see when you fine tune it on that many images
https://twitter.com/novelaiofficial/status/1582834597846339584
I'm struggling on making the training script run properly with that much images
it always automatically shutdown and left no logs or errors.
@kblueleaf interesting, keep us updated though I guess
脸上都带红晕是什么奇怪的模因
新的版本有什麽區別嘛，我試了一下，似乎沒有什麽區別。
沒有太大的區別
但是增加了未指定內容的多樣性
（例如，沒有指定顏色的前提下更容易出現不同顏色的飾品等）
@kblueleaf 好噠，感謝青葉大佬。
话说这个有Refiner吗，相同分辨率普通模型一分钟换成xl出图就要十几分钟了呜
這個不用refiner
然後你遇到的問題是爆顯存
my picture becomes really noisy after process, (i uploaded it at the bottom for reference). where did i do wrong?
Please use vae:None to utilize the var built in the model
Or choose sdxl's vae
@kblueleaf i see, thank you it worked
Can someone unironically help me with sdxl? Every time i try it i get aweful ass images that look worse than even some 1.5 images.
maybe its your res, SDXL uses 1024x1024, and 1.x uses 512x512, and the VAE, you have to use an SDXL
VAE
@RyuGure Maybe? Couldve also been just me trying to run it on colab because i didnt have a good pc at the time
sometime it output 2.5D style image and hard to control