More detailed tutorial or video

[axsddlr]

would love some more info on how to utilize this

[ljleb]

Not sure how useful this will be, but here I go anyways. I was pondering on what some of the buttons do, so I looked at the code and tried to figure some of them. I'm still in the process of understanding the rest, but in the meantime here's 2-3 things I found:

• prompt 2 and negative prompt 2 do nothing if you leave them empty. the extension puts images generated from prompt1+negative1 and prompt2+negative2 together after generating 2 different sets of images. basically the purpose is to be able to generate different kinds of images to train the model on, instead of only a single type of image. I suppose picking prompts as different as possible for prompt 2 and negative 2 will yield more general merges (i.e. merges that work with more prompts):

https://github.com/Xerxemi/sdweb-auto-MBW/blob/82fc57f92697ed20878da710e7f5b4361eca1676/scripts/mbw/auto_mbw.py#L428-L433

• "multi pass/merge progressive" checkboxes increment the seed on some occasions. I don't understand the full context but that's a start. They don't have any effect when chk_keep_random_seed (Keep random seed) is checked

• Tally types 2 and 3 only useful when passes > singlepass. Tally type 2 only useful when passes >= doublepass. Tally type 3 only useful when passes = triplepass. A tally type is how they combine the predicted aesthetic score of multiple images for a single pass

[Xerxemi]

what is said above is correct. I'm pretty busy currently so getting around to writing official documentation would take a while.

The double prompt was added as a method to tune on two different variables and eliminate prompt dependency, similar to the progressives which were added to reduce seed dependency (the current progressive has some issues however).

[ljleb]

I won't be able to understand everything I think, but I can put a short document together for everything I end up making sense of. If you could simply verify that the info is correct, or maybe just fill 1 or 2 gaps, that would be a wonderful contribution for sd enthousiasts I think. I totally understand being busy, I'm just suggesting solutions.

This repo is pretty useful, you can probably use the receipes it finds as a good starting point, if not even as an acceptable destination for a merge. (currently in the process of running my first auto merge so i don't know exactly how effective it is)

[Xerxemi]

Contribution would be helpful. Though, some of the documentation requires subjective general recommendations which I'd probably still have to add.

Results do depend on the settings, and as the creator I do think it is quite effective (obviously lol). But currently, the defaults don't really make much sense tbh. Documentation would definitely be useful for this.

It can definitely be used as either a general merger or an baseline to tweak, I've seen cases of both. You can also get it to tune only certain layers via the checkboxes.

EDIT: sorry for the excessive editing btw, it's just a habit lol.

[Deph1]

I think just a general guide on how to get it up and running would be useful with an idea what your preset would look like.

Personally finding it difficult to even get it to run right now. Zero errors but the wiki isn't helpful on how to get it going.

I appreciate that as a relatively new individual to github I'm at a disadvantage already but I thought I would chime in from this perspective as the tool is meant to help cut down on the technical aspects.

Appreciate you making such a tool and the outputs you've created look fantastic.