How is the pos/neg weight set?

[ZichengDuan]

Hi, I appreciate your work, and your paper is well written and easy to understand, but there's one issue I haven't found the answer for. I have read through the code and paper but still have no clue about how the weight of positive/negative weight is set. I can see that they depend on the time steps and classifier-free guidance, but still don't know the details. Can you please provide some insights regarding this? Much appreciated.

[dvruette]

Hi, thanks for the kind words! The weight of positive feedback is basically a scaled and shifted square function, with the min_weight usually being set to 0 or a small value like 0.1. Here's an illustration:

The feedback strength for negative feedback is set as a fraction of the positive weight, usually 0.5. So the shape is the same, it's just the scale that's different.

Looking at the code here, it really is quite convoluted (and also implements some additional features), but the SD webui extension is much cleaner and IMO easier to understand: https://github.com/dvruette/sd-webui-fabric/blob/main/scripts/patching.py#L56

I hope I was able to answer your question, lmk if there's anything that's unclear still!