Disclaimer
This model applies Terminal SNR to its training noise schedule, as described here.
This will not work out-of-the-box (as of June 2023) with Automatic1111's Stable Diffusion WebUI.
It works great using a Diffusers pipeline, such as InvokeAI, or a simple Diffusers python script using DDPMScheduler. One such script is available, here.
I hope that Automatic1111's support for terminal SNR pipelines is improved someday, because SDXL will require it.
Description
This is a checkpoint trained from MJ 5.1 data from a random sampling over a single day's worth of upscaled images.
It was started from stabilityai/stable-diffusion-2-1 with a mix of 22,000 synthetic images and about 7,000 real photos.
MJ data was captioned by its users. Real photos were captioned by BLIP using the Salesforce model on top of the SD 2.1 text encoder.
Some helpful terms:
UHD
8K
Sharp
best quality
midjourney / midjourney's
Here are the most frequent terms in the training data captions:
in: 7348
and: 6337
with: 6110
style: 3331
on: 3097
white: 2643
background: 2380
black: 2289
realistic: 1937
art: 1589
an: 1422
.png: 1344
beautiful: 1313
full: 1228
light: 1200
is: 1192
hair: 1192
detailed: 1189
woman: 1109
from: 1108
for: 1075
logo: 1013
wearing: 1013
body: 1012
dark: 1004
to: 998
high: 987
blue: 949
beta: 933
8k: 928
illustration: 915
man: 915
red: 904
cute: 889
photo: 873
at: 857
as: 849
design: 837
relaxed.png: 787
cartoon: 761
colors: 745
old: 714
portrait: 709
ultra: 692
fantasy: 682
eyes: 680
green: 678
4k: 667