New Node: Token Strength over Time #9707
wynandhuizinga
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
I have a proposal for a node:
Token Strength over Time:
I'm currently experimenting with strength of tokens over time. I built a nightmare of a node network to pull it off and it does look promising and pull out way more details out of images. Philosofically speaking, If you try to make a picture of a landscape with a bird in the sky... it's relatively easy to generate a blue sky if already some parts of the latent image have reverse diffused to blue. For a full noise image however (at the start), you want to emphasize large chunks need to become sky. For the bird in the sky however, you want the opposite: In the beginning, you don't care about the bird yet, That only becomes interesting nearing the end. And why would you let it take attention away at the start.
The idea is simple: get to a state were EVERY step of the (rev) diffusion process has its own 'per token weighted' prompt. I'm no great UX'er, but as per the proposed mock-up, can come in either 2 flavors. With dials or with Graphs. Graphs would be nicer, but probably more tedious to implement. Alternatively, you can also imagine having 1 big graph with multiple lines in there where each line represents a token. Probably too tedious to keep it workable though.
As for the background, currently working with Illustrious which doesn't seem to accept this style of prompting --> [ ... : ... : 0.4 ]
Even if it did though, your prompt would become very unreadable, besides, it only allows for 1 switchover moment.
Very curious to hear your thoughts about it.
Beta Was this translation helpful? Give feedback.
All reactions