Quick Addendum/Repost to my original merge Post last thread.
>What the fuck is a Bayesian Merge? And why the hell are you so excited about it?Well, it's quite simple, Bayesian Merging takes everything that is so great about MBW Merging and does it for you. And unlike other Auto MBW mergers, this isn't a brute force algorithm, but rather does Black Box Optimization on the merge. When combined with a scorer, you can achieve results that would have taken days to produce and test in only a couple hours. And I think the results speak for themselves.
>Well that sounds great and all, but I produce perfectly fine results with standard merging.Well, MBW Merging is for merging two models without losing/diluting the concepts from either. The BBW model is an almost perfect candidate for MBW Merging, because with any sort of standard merge, you lose/water down the main concept of the model due to the fact that fat anatomy is so rare in the NAI dataset, and in general there is a very small dataset of art for the concept, which makes it easily overwhelmed when merged with another model. The fact that I didn't have to screw around with individual Block weights and continually test everything manually is a massive plus, since it saves time.
All in all, Bayesian Merging was absolutely worth the trouble it took to fix (it's now fixed on main as well so you should be able to use it), and configure. I'm curious to hear some feedback on how it performs since I went to some pretty big lengths to make a pretty large payload set to cover a number of holo chuubas.
Here's the original post:
I finished the final version of the bbw_hll3 merge.
Test Images from all 13 payloads. (Some of these have Dynamic CFG, some of these don't, all of them have been through Hires fix. These prompts were all run on 20 steps DPM++ 2M Karras( with most having 7.5 cfg) when scored) There is 1 non-fat prompt, where I just prompted for a character with no modifiers, and one prompt that should result in an image that isn't fat focused. (Polka and Noel)
Test Images from all 13 payloads. (Some of these have Dynamic CFG, some of these don't, all of them have been through Hires fix. These prompts were all run on 20 steps DPM++ 2M Karras( with most having 7.5 cfg) when scored) There is 1 non-fat prompt, where I just prompted for a character with no modifiers, and one prompt that should result in an image that isn't fat focused.(Polka and Noel) Additionally, both the OP Image and the request Anchor were generated from this model.
Suisei:
https://files.catbox.moe/izd9b9.png Watame:
https://files.catbox.moe/zj0uj7.png Polka 1:
https://files.catbox.moe/9yxuo0.png Polka 2:
https://files.catbox.moe/i8v713.png Polka 3:
https://files.catbox.moe/km66fr.png Noel:
https://files.catbox.moe/7ari39.png Mumei:
https://files.catbox.moe/jo92in.png Mio:
https://files.catbox.moe/cloew1.png Lui 1:
https://files.catbox.moe/bjeein.png Lui 2:
https://files.catbox.moe/u6xa1y.png Lui 3:
https://files.catbox.moe/poink3.png Lui 4:
https://files.catbox.moe/llvbl9.png Laplus:
https://litter.catbox.moe/gi73mt.png Kronii 1:
https://files.catbox.moe/gfqfo1.png Kronii 2:
https://files.catbox.moe/3ci6lt.png (using DPM++ 2s a Karras at 41 steps)
Irys 1:
https://files.catbox.moe/gnrtrb.png Irys 2:
https://files.catbox.moe/dqnob8.png Irys 3:
https://files.catbox.moe/aaqmi1.png Ina:
https://files.catbox.moe/80rcfc.png Haato 1:
https://files.catbox.moe/e5k7is.png Haato 2:
https://files.catbox.moe/1unn0s.png (detailed background added to tags)
Ayame:
https://files.catbox.moe/pknk3a.png Model Download is here:
https://huggingface.co/mariaWitch/ExperimentalBiggerMerges/resolve/main/bbw_hll3.1%20Gods%20Amongst%20Men%20Edition.safetensors 