New thick merge from me. Went a little bit much towards thickness, but i think it turned out quite good. For you to decide anyway, have fun! No need for tricky prompt to get good picture — what you see is what you get — feel free to experiment.
Two versions finally — with baked in and without VAE (i use sdxl_vae). Hope this baking works — please, let me know.
Low CFGs and face fix for small faces.
Description
FAQ
Comments (17)
Is there a reason why you don't just bake the VAE like almost everyone else?
Yes, the reason is simple - have no idea how to do this. I use Supermerger in A1111. If you know how to bake VAE there - let me know. Thanks!
@Ungebliben I don't use A1111 so I can't check for myself but according to a quick Google search, there is an option that let's you chose a VAE to bake into the merged model.
@Lizardon1025 exactly THAT flag I'm being searching for since I've started getting this type of comments. No progress since. That's citation from Supermerger Github:
"Bake in VAE. When saving the model, the selected VAE is incorporated into the model." That's it. No options, no flags, nothing regarding VAE in the interface (including A1111 settings), even no dropdown\option with VAE selection as mentioned in the quote. I'm ready even to bake it separately, if it's possible, and you can give me the tool. Problem is that i firmly stick to A1111 because it's easy, logical, straightforward and works right out of the box. Everything else just the waste of time for some funny pictures (with all respect - I'm not going to setup a environment for 5 minutes to render each 1-2 images). A1111 is what it is — automatic indeed. But with the cost of fine tuning and tricky options. Anyway, I'm open for suggestions about baking vae.
@Ungebliben I understand that you want to stay with Automatic1111 and that's perfectly fine. In a Youtube video called "Automatic1111: Super Merger Mini Tutorial (Remixed)" by Duskfall at 1:38 I found the option to "Bake in VAE" under Options. It seems to be quite hidden but in case the layout hasn't changed to much since the video was created, it might still be there.
@Lizardon1025 Thank you, i'll give it a try!
For a1111 fan, the Forge extension may be of interest?
https://github.com/lllyasviel/stable-diffusion-webui-forge
If you take a lot of pictures, you can use swarmUI to process them on multiple computers. If you have a lot of checkpoints and LoRas, you can create presets for them and call them up with just two clicks. https://github.com/mcmonkeyprojects/SwarmUI
...I know this wasn't asked for, but I feel that there are readers who need this information. Good luck in the world of images.
@Sandi22 Thank's a lot, i might give it a try as well.
I've tried to bake VAE in, let me know if it worked, please.
I've been getting into this a little, and one thing I do is, run it back through the default checkpoint merger in A1111. I put the weight at 0 and run it through with my desired VAE. Maybe that's a bad method, and I'm crazy for thinking it works. I don't know. I'm still pretty new to this, like I said, but it seem to do the trick.
@Ungebliben Baking the VAE worked perfectly, thank you :)
https://civitai.com/images/102515078
@lizardon1024 should I publish models only with baked-in VAE in future, in your opinion?
@Ungebliben I guess so, otherwise people with little experience will try out this model and simply discard it once they get the usual artefacts of a missing VAE (washed out colors & white spots). I don't see any reason why not to bake a VAE ^^
Creating some images based on checkpoint but images are generated foggy and with some white dots.
Utiliza algún VAE. Cuando salen esos puntitos blancos y la imagen como apagada, es porque no hay un VAE en uso.
Check if your prompt is controversial (pos don't repeat neg in any tag), check CFG - it chould be low (3-5). If using lora try rendreing without and see if it helps. Maybe it's vae, try version with VAE baked-in, it's uploaded here. I render at 30 steps, try this as well. Try prompt and seed from one of my images and render with same settings for a start.
@MirandoPaCuenca Gracias!!