v2:REBIRTH

1.2 was incorrect upload. 1.2.1 is the correct one. delete 1.2

@s6yx thanks

yes, bayesian merging technique only merges UNET values. That removes a lot of the unneeded weights

up to you. 1.2.1 is an improvement of 1.1

@s6yx  Thank you, Well if its an improvement and smaller ill replace it then! :) Great work on this one!

give me a few and i can make a version for invokeai and put it on huggingface

invokeai/revAnimated_v121_invokeai.safetensors

@s6yx Thank you, I use this model in one of my mix's and was trying to update the Model =)
You do great work

@s6yx Generally there shouldn't need to be a different version for Invoke. Any idea what is going on here?

@void2258 https://civitai.com/models/5585/deliberate-for-invoke

@void2258 I was wondering that myself, as I also use InvokeAI and could not for the life of me get the RPG model to work.

@s6yx Thanks for posting the InvokeAI version! Going to download it and try it now !

@earlyapriladam i would hold off on that. working on 1.2.2 release to fix somethings

@s6yx 1.2.2 working nicely for me in InvokeAI, thanks for the update 🙏

@lenymo no problem

Ok. I've redownload model and it works again)

I've trained now around 8 on V1.1 and i don't think that i set something special in kohya.

many people would say otherwise. redownload the file and restart auto1111

Breaks in Comfy UI as well without a special workaround.

@void2258 i dont use anything else besides a1111 so again. I would not know

@s6yx Spoke to people at Invoke and Comfy. Missing CLIP weights, and without them it won't work in Invoke at all. It can be forced to work in Comfy with a performance hit (though no guarantee 'work correctly', just won't immediately fail). Also apparently doesn't work in all forks of A1111, because it's counting on something the main A1111 is actually doing wrong to function.

@void2258 message me on discord: S6yx#3618

@s6yx What discord server? Also, I am just relaying what people told me, so if you want technical details, probably need to go the invoke or comfy discords.
A bunch of people are saying lora training won't work. This is probably related to the same problem, since lora plays with the information that's missing.

@void2258 working on 1.2.2, with a different merge method. Seems the bayesian merger technique just keeps UNET Model component, hence why smaller model size

uploaded

Do you guys know if I can use the inpainting version for general use? My reasoning is that it may have more precision or be less pruned...

You guys are awesome! Thanks!