[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/vt/ - Virtual Youtubers

Search:


View post   

>> No.53482822 [View]
File: 2.15 MB, 1600x823, cosineA.png [View same] [iqdb] [saucenao] [google]
53482822

>>53480859
Basic Tier Addendum:
One thing I forgot to mention, with this level of merging you are hard capped a the amount of models that can be merged together before hard burn in starts to appear. So it's best to be careful, and not go to overboard as you may end up going backward.
>>53482475
Yes. You can keep terrain and backgrounds similar to one model but the pose and artstyle from another. Pic related is a good example of this. (this is a merge of FeverDream and AnythingV3 on CosineA merge mode.) If you want to go super specific we are going to get into Block Weight merging which will be one of my next posts.

>> No.52694324 [View]
File: 2.15 MB, 1600x823, file.png [View same] [iqdb] [saucenao] [google]
52694324

>>52680966
>I see, is there a reason to choose 2 models that are similar? I sorted some of my models, they seem to fall into 3ish categories. Pastel/anime, toon and real. I haven't done much with them to be honest but would one expect to get some wild and interesting mixes between the categories? Or maybe most will become a mess.
>>52683686
>Makes sense, but some times it can feel like spending a ton time mixing something that ultimately if not put side by side you can't really tell it's a new model. I'll still try and mix different categories someday for experiment sake.

Both of the above points are correct, depending on the type of merging method that you use. Certain Merging methods are significantly better at merging style/structure of two models. (Cosine Merging comes to mind. See Pic related, which is a CosineA merge results between FeverDream and AnythingV3) There are certain other merging methods that literally "train" the merge output of each block. And then you have the Bayesian Merging method, where you literally can either have a scoring model do scoring for you in order to try and find the best model merge, or you can score all the images yourself to optimize the model to what you believe is good. As someone who HAS done this, this is incredibly fruitful at producing a merge that both creates good looking images but keeps concepts from one model while keeping the structure of another. BBW Aurora Borealis is my work, and that I think is the best model so far in terms of pure image while still being able to do fats. Although that has it's own downsides, as you have to score 1000+ images in order for the merger to be able to produce something of value.

At the end of the day Block merging is very much black magic unless you do indepth research into the effects on the merge of each individual block. Every Model is different, and so each block is going to contain different "content". My general advice is to ditch the default SD Webui model merger, it's honestly garbage for anything beyond basic merges. Supermerger is the BEST merging tool (that doesn't automate the merging process) by far.

>>52693467
>BBW BBWM Anon's
Hey that's me!
>DoritosLocosTacos&IRyS
Kek I like the new name

Navigation
View posts[+24][+48][+96]