IMPORTANT: to get the best results with this model and prevent oversaturation, you must use "guidance limiter" (included in reforge, comfy node) with it's default settings.

information

noobai vpred merge, many models were used, but that is the base model as target.

the only non-illustrious based model merged is lustify (nsfw) it's weight difference was added via perpendicular component a little to help character positioning/backgrounds

merge method

sd-mecha via comfy-mecha using many of it's methods like tensor sum, rotate, multiply quotient, geometric median etc. and a modified karcher-mean script. (original source)

(the diff from the source is that all tensors are converted to a float type for the calculation and then it converts the final result back to the original data type to maintain model integrity.)

recommended settings

interface: comfyui or reforge

samplers: res_multistep or euler (and their cfg++ variants)

schedulers: either beta or ays+ (alignyoursteps but with sigma set at 0, it's available here)

step count: around 20

cfg: 5 to 6

prompts: include "very awa", any danbooru/gelbooru/e621 tags. (some natural language also works)

negative prompt: low quality, worst quality, etc.