This is my first attempt at blending, still testing, and I don't understand model training at all, so I don't guarantee quality. It's basically just for fun.
2023/10/13
V0.11 fixed the v0.1 version's problem that can't work well without VAE,
Just use the V0.11 + Hires. fix, it can work without using VAE.But after a short test, even if you use the VAE, it seems that it won't have too much effect to images,need more test.
2023/10/13
V0.1 will be very trash when used without VAE,
(V0 can work without VAE.)
VAE I USE:
https://huggingface.co/JamesFlare/pastel-mix/blob/main/pastel-waifu-diffusion.vae.pt
you can also try other Vae and maybe find a better one for yourself:
Anything / Kl-f8-anime2 / Vae-ft-mse-840000-ema-pruned / Blessed / ClearVAE (fp16/cleaned) - anything | Stable Diffusion VAE | Civitai
VAEs in 2D/anime/art/illustration genres should all work...maybe.
Mix by ...three models maybe ?
If i didn't doing wrong :
Bomain 64 - v1.0 FP16 no ema | Stable Diffusion Checkpoint | Civitai
AnyLoRA - Checkpoint - bakedVae (blessed) fp16 NOT-PRUNED | Stable Diffusion Checkpoint | Civitai
Pastel-Mix [Stylized Anime Model] - Pastel-Mix [Pruned FP16] | Stable Diffusion Checkpoint | Civitai
already played a little bit, and all were better than I thought when doing the color pictures, detail eyes, and the tooth generate,so I'll take it...for now.
This mixed checkpoint is good at anime, illustration, sometimes 2.5D.
NSFW can be generated too.
If the model has many problems, controversies, or just too bad, it will be deleted directly.
Description
The generation quality of this model seems to drop off dramatically when using Regional Prompter, but I'm not sure if this is just my personal situation or else.
This version also doesn't require VAE.
FAQ
Comments (3)
Perhaps a stupid question, but I wonder: Why is this checkpoint heavier than the others?
...No, that's actually a good question,because I wanna know too.
I've tried a few things regarding this, but until now, I can only vaguely suspect it has to do with the size of the model I initially used for blending like AnyLora Full Model fp16 (3.22 GB) and Bomain 64 v1.0 Pruned Model fp16 (3.59 GB), causing the file to remain bulky even after "Save as float16".
The only solution I can currently think of is to restart the blending process using a Pruned Model of 2G or less and try to ensure that the styles remain similar.
However, I'm still hesitant about whether to proceed this way, and I'm not sure if I'll be able to generate something similar or better again, especially considering that this checkpoint was born out of playfulness and accident from the beginning.
if I manage to create models of reasonable size and I like it, I'll be sure to share them. However, it's likely they'll be released as a new separate checkpoints.
Thank you for your input and inquiry,hope this answers your question, although I must admit my own understanding is limited.
@Junbegun Thank you for your frank and sincere response!
Know that I tested your checkpoint and that I continue to test it and in all sincerity, I really like it!
Thank you for your work and I hope you will have other beautiful creations to offer! <3