Haoming02
Implement dynamic weight formula in Advanced mode
This provides the option to use dynamic mask weight to more smoothly blend the conditioning; for example, you can set the weight as =exp(-((x-0.5)**2 + (y-0.5)**2) / (2 * 0.15**2)) to give it a Gaussian distribution (bell durve) that falls off smoothly as it moves away from the center of the image, or use two regions with respective weights =y and =(1-y) to gradually transition between the prompts as you move from top to bottom along the image.
Since the effect of this Extension isn't perfectly precise to begin with, doing these fancy maths feels rather... redundant.
Do you have any examples where these changes improve the outcome substantially?
Hi @Haoming02 , thanks for this outstanding extension, and thanks for the very quick feedback on my PR -- sorry I didn't mark it as a draft at first and provide more context just yet (I'm at the end of a long day here and I forgot that you would probably see it right away due to the time zone)
Since the effect of this Extension isn't perfectly precise to begin with, doing these fancy maths feels rather... redundant.
Yeah, I agree the Gaussian may be a bit overkill, but the linear interpolation works really nicely... I'd suggest to give a try to create a transition with styles, texture, mood, background etc, and still use the same sharply-bounded boxes for precise local character control.
What I propose to do is this:
- I'd like to add a 't' variable for time step / noise level, this would be extremely useful as you can control which prompt applies to more coarse or fine details of the image and conserve a lot of context this way. For example, by not applying a style prompt before it's actually needed, and dropping out a composition prompt once the base composition is set and the model's self-attention can maintain it.
- I have a list of more dynamic weight formulas to test and would like to add the good ones to docs/examples and consider adding the most useful ones as presets to make it easy for users to apply without any fancy maths
But I need to sleep first as it's already very late here. :P