UPDATE: safetensor(float 16) and Pruned version file for PastelMixAlike
New Version: PastelMixSam
This version has SamDoesArts LoRa embedded in it.
NOTE: If you're using PNG info from my old sample(Delete the old hash from "Override settings" as it disables the embedded LoRa and uses only the model.
UPDATE: After testing, I have concluded that PastelMixSam yield the same result as PastelMixAlike. Just remember to delete the hashes when copying model hashes from PNG info and you're all good.
And Trigger Word is not necessary for both versions. The result will change, for the better or worse (They all look good)
A new version is coming that will enhance the THICC and sexiness of the body. Stay tuned!
PastelMixAlike
Just my fine-tuned version of PastelMix
No vae is needed but nai.vae is the only vae that i think that are somewhat decent here.
Clip skip=2 and "quantization in K samplers" is enabled
Several Lora was used to test its compatibility:
SamDoesArts (Sam Yang) Style LoRA as the Base Lora:
https://civarchive.com/models/6638/samdoesarts-sam-yang-style-lora
Compatible well with these Lora: (Need some prompt tuning)
Recommended settings are included within Sample images
Currently, there are purple artifacts that are affecting images. If anybody has a fix please comment down below.
Description
Currently, there are little purple artifacts that are affecting images. If anybody has a fix please comment down below. Editing them out is not hard however
FAQ
Comments (17)
Purple spots (and other color artifacts) are a common result of not using a vae or using a vae that doesn't quite fit the model. Try using a few different vae's and see if that helps
Is there a vae that would remove the purple spot without having a minimal effect on the image? I don't think such vae exists
@DegenerateAkko All the vae's are like touch-up models. Anime vae's work well with anime models and realistic vae's work well with more realistic models, but they're just meant to fix small details like the problem you're having. They don't drastically alter images, just small changes here and there. I don't really know all of them, but off the top of my head there's pastel mix vae, anything v3 vae, kl-f8 vae, (something)-mse 840000 and more
@Bunger_Beesechurger Already tried, kl-f8 vae, (something)-mse 840000, not my fav. They alternated the art style by changing the colors and stuff. I will try anything v3 and pastel mix vae. But realistically I'm just gonna try to edit the artifact out. Any tips?
@htke Yep, any idea how to reduce file size?
@DegenerateAkko as simple method in the merge tab(built-in) check the box *save as flout16* or whatever
there you can also backed VAE, that's very useful feature @DegenerateAkko
@DegenerateAkko you can also install the model converter addon from the local install extensions menu to prune the model.
@OsirianArts What Pruning Methods and Precision should I use?
@DegenerateAkko no-ema and fp16
is there any chance on reducing file size??
For now, No. If you're using it with google collab, download it on the notebook collab instead of google drive. I will figure something out
if you are using automatic11111's webui, Go to Extensions -> Install Model Converter -> Restart UI -> Go to Model Converter tab -> choose the model -> Precision: fp16 ✓ and Pruning Method: no-ema ✓ -> Run
At least, it should decrease model's size to 4gb i guess.
i love this so much!more common-use checkpoint plZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZ!!!
This is Clickbait..
Why don't you just Merge your model with Samdoesart lora..
Also throw some vae into it..
Sr, I already have it but the merge version was buggy and didn't produce the LoRa result. Then I realized my Prompt from an old PNG info image included its old hash and it somehow disable the embedded LoRa. I would still recommend using Them separately though ( less problem to deal with and more flexibility)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.