lucidrains / deep-daze

Simple command line tool for text to image generation using OpenAI's CLIP and Siren (Implicit neural representation network). Technique was originally created by https://twitter.com/advadnoun
MIT License
4.37k stars 327 forks source link

create_story feature generates washed out images #60

Closed afiaka87 closed 3 years ago

afiaka87 commented 3 years ago

I'm seeing lots of commented out normalizations. Any reason in particular those need to be gone? I noticed this feature wasn't exactly advertised anywhere (and still isn't accessible from the cli). Does that mean it's not complete?

NotNANtoN commented 3 years ago

Hi! The normalizations were faulty indeed. You're right, the images look more washed-out. I think that's because it is actually correct now. Before, the (unnecessary) normalization turned the images much darker and distorted the colors. That leads the image to be less washed-out (darker) and distorts the colors, but it is simply not correct (could be compared to some Instagram filter).

I fixed the scheduling system in #61 and if you use this git head you will see much better results (less washed-out) than before.

NotNANtoN commented 3 years ago

You're right, the story creation was not yet advertised by me, because I feel like it's still a bit experimental. In the new PR you can see some examples for it.