@WAI0731 oh i see, so it comes from the original model creator. Thank for answering the question so fast!

"Very Awa" = Very aesthetic , it's equivalent to Pony's score_9 as the top x% images based on an Aesthetic Scorer received the tag

@Churrus Ooooooh!!, thank you!

That is the aesthetic label obtained from the aesthetic scoring model. So far, there are only two: "very awa" and "worst aesthetic." The former refers to data in the top 5% weighted scores from waifu-scorer-v3 and waifu-scorer-v4-beta, while the latter refers to data in the bottom 5%. It is named "very awa" because its aesthetic criteria are similar to the (A)rti(Wa)ifu Diffusion model.

@0xSeiunSky Thank you!, now im cristal clear! :D

Have you tried changing the way you prompt?

@madaraxuchiha88 I'll try some more methods, but so far my experience with V-pred isn't great

The success of NovelAI3 might have led people to favor V-prediction, and indeed, V-prediction has advantages in color and instruction adherence. However, considering the inadequacy of infrastructure, the benefits of using V-prediction remain unclear

Eps is a flawed method, it can't produce completely bright or dark images and always try to bring the average brightness of images to 50%, making it add/remove/change color of elements unprompted to try to satisfy that. Vpred fixes that and has superior prompt adherence because of it.

Try to make a night or very dimly lit scene in a non-vpred model and you will see various lanterns and light sources sprinkled throughout the image. Now you know why.

@Churrus v-pred might be mathematically better, but like the others say, I've yet to see a v-pred model that outputs results that are even subjectively nicer looking than with ε-pred, let alone objectively. Higher dynamic range? Sure, and for a particular type of image (stylized high contrast 'artsy' kinds), it can look pretty decent, but the severe drop in background quality and style adherence far outweigh any benefits for general use imo. I also haven't seen any of the improvement in prompt adherence during actual use (though it's never been something I've struggled with so much on ε-pred to begin with).
No idea if these are just flaws with NoobAI specifically though rather than v-pred as a whole, but until a decent v-pred base model comes along I think I'll be happily sticking with ε.

This v-pred01 and v2.0 are totally different.

Use it and you'll see.

@WAI0731 v-pred rocks its skill issue what separates the great artists

V 比较“听话”，比如 step on a car 这个动作在E的noob或者其他模型上怎么都出不来，但是在V上很容易就出了

@brahianvalles bruh you aint a artist, just shut up

@WAI0731 I changed the scheduler but it didn't work properly. I tried the same prompt as the sample Satono Diamond illustration, but this also caused color saturation when the CGF scale was 6 or higher.

Sample: Euler a I ran multiple schedulers and they all saturated.　hmmm

@WAI0731 reForge

@WAI0731 I found the cause.

With Kohya HRFix Integrated on, color saturation occurred when the CGF scale was 6 or higher.

This is a Reforge feature, so it's not an issue with the model.

However, it doesn't happen with other models, and this is the first time I've seen this issue.

is it better than other illustirous merges? And in what way

@WAI0731 I set the CFG Scale to 5 or less and for some reason all the problems went away.

Thank you very much.

if you don't mind me asking, what is this hyper lora?

@WAI0731 大佬，请问你是会把这个大模型投到illustrious分区还是NoobAI分区

（lib把illustrious跟NoobAI分成了两个分区，而且LORA还不通用，很烦）

@WAI0731 请问大佬有没有能力跟lib官方沟通，illustrious跟NoobAI的Lora不互通真的很难受

@WAI0731好的大佬，很期待这个大模型的表现

@WAI0731 哦对了大佬，把ponyXL也更新一下吧，谢谢

你觉得的模型各自的优劣哈哈哈~

I think the answer is yes :D

Thanks for this! Do you have any intention to do this with the v-pred-2.0 version as well?

@bluehands02 Yes, I will convert to the GGUF version of the model later

@bluehands02 

They will be released here:
https://huggingface.co/btaskel/wai-shuffle-noob-vpred20-GGUF

https://huggingface.co/btaskel/wai-shuffle-noob-v20-GGUF

@Bt_Askel Thanks!

How exactly can a quantized version achieve better quality?

@_Y_ For SDXL, FP16 is unnecessary, and Q8 is usually sufficient to achieve the same quality as FP16. For models that can run on devices with sufficient VRAM, I recommend choosing the quantization method without the K suffix, because it offers higher performance. For devices with insufficient VRAM, you can try the quantization method with the K suffix, which can achieve higher quality than the non-K method.