all rights and congratulations to: NoCrypt https://huggingface.co/NoCrypt/SomethingV2_2

SomethingV2.2

Welcome to SomethingV2.2 - an improved anime latent diffusion model from SomethingV2

A lot of things are being discovered lately, such as a way to merge model using mbw automatically, offset noise to get much darker result, and even VAE tuning. This model is intended to use all of those features as the improvements, here's some improvements that have been made:

Can't trust the numbers? Here's some proof

Recommended settings

VAE: None (Baked in model, blessed2)
Clip Skip: 2
Sampler: DPM++ 2M Karras
CFG Scale: 7 ± 5
Recommended Positive Prompt: masterpiece, best quality, negative space, (bioluminescence:1.2), darkness, dark background
Recommended Negative Prompt: EasyNegative
For better results, using hires fix is a must.
Hires upscaler: Latent (any variant, such as nearest-exact)

Recipe

Due to SD-Silicon's Terms of use. I must specify how the model was made

Model AModel BInterpolation MethodWeightNamedpepmkmpsilicon29-darkMBWReverse CosinedpepsilisomethingV2_1dpepsiliMBWCosineSomethingV2_2 rawSomethingV2_2 rawBlessed2 VAEBake VAE-SomethingV2_2

Why not call it SomethingV4?

Since this model was based on SomethingV2 and there's not THAT much of improvements in some condition. Calling it V4 is just not right at the moment 😅