When you generate images with VQGAN + CLIP, the image quality dramatically improves if you add "unreal engine" to your prompt. People are now calling this "unreal engine trick" lol e.g. "the angel of air. unreal engine"

May 31, 2021 · 9:02 PM UTC

Replying to @arankomatsuzaki
Has nobody backpropagated to find an optimal general prompt shibboleth yet? A Good Loss Function For Quality Is All You Need 🙃
"Could You Please" (as a prompt prefix) is All You Need
Replying to @arankomatsuzaki
Can it generate comic style drawings? E.g when wall-e came out, there were articles where it draw "perfect" cartoon styled avocados etc
We didn't try enough number of comic style drawings to succeed, but we succeeded with colored pencile style, which may be of your interest:
Replying to @arankomatsuzaki
nice! can I test that myself? is there any colab or description of how to set up the VQGAN+CLIP model?
I found some relevant links here, not necessarily the same as what I'm using tho. Maybe you can find more info in the same subreddit: teddit.net/r/deepdream/comme…
Replying to @arankomatsuzaki
Hi, I thought this was cool, but I took this link, entered your description text, and got the following result. I want to know your parameter Settings, seed, model, size, etc colab.research.google.com/dr…
This is an image taken from EleutherAI's discord, and someone else generated this image. So, unfortunately I don't have the info on that :(
Replying to @arankomatsuzaki
Can you try “the angel of air. minecraft”?
It didn't perform as well, unfortunately, but "the Nether from Minecraft rendered in Unreal Engine" looks nice: