sdweb-auto-MBW VS stable-diffusion-webui-model-toolkit

Repo: https://github.com/Xerxemi/sdweb-auto-MBW

AutoMBW is an extension made by Xynon#7407 and Xerxemi#6423, and performs block weight merging automatically.

Alternatively, something I've been working on is applying block merges to the fantasy art style model. However, even with auto block merging (https://github.com/Xerxemi/sdweb-auto-MBW), I can't get anything decent. I probably don't understand enough of the block merge theory, yet.

I haven't tried this out yet myself, but it's something that looks major. If I read the article correctly, autoBWM basically fiddles with the weight merge sliders and puts the result through aesthetic scoring in order to discover the best results (according to the aesthetic scoring thing(s)).

I think you can prune your model to under 4G using the webui-model-toolkit

I am using Model Toolkit in A1111, and most of the time it does it's job well by removing junk data and converting model to fp16. This results in 2GB file, but... There are models that cannot be converted to 2GB version. Noticeably CyberRealistic v2.0 has this "issue".

Use the model toolkit extension in A1111 -> https://github.com/arenatemp/stable-diffusion-webui-model-toolkit

I merged this on Colab, so you don't need a powerful graphics card. You'll need the block merge extension,the model toolkit extension and the supermerger extension.

If issue persists after trying the suggestions given, fix the model by using the model toolkit: https://github.com/arenatemp/stable-diffusion-webui-model-toolkit

Also, I'd highly recommend getting the Model Toolkit. It's an extension that can load up your models and prune information it doesn't need, making them drastically smaller. It can also fix broken models. You have to install it manually, but it's pretty easy (and explained on the GitHub). Well worth downloading.

Protogen Infinity has been my go to in recent weeks (I recommend getting the full unpruned version, and pruning/fixing it with arenatemp's toolkit). Very versatile, flexible, creative, diverse, high quality, etc., but I think I'll be trying Realistic Vision 1.4 (and maybe Deliberate v2) more now. I might try merging Realistic Vision and Protogen Infinity, and see if that improves on either of them (merging RV and Deliberate together didn't necessarily see any improvements).

Someone made a "model fixer" for this reason too apparently.

I agree about having to reload the UI. Since I have some pluigins installed I have to restart the UI or the engine (via CTRL-C) to maintain that 576x576 resolution as it'll run out of just enough memory to do it sometimes. I've found that safetensor format makes reloading MUCH QUICKER after a restart and I've used the Basic tab on this extension to trim the fat out of models I use, while converting them to fp16 and safetensor format.