This publication is a testing ground for recreating the ideal model. These versions of the models are additionally trained on additional datasets on yutamix pxl. Since they are very different in appearance from the original, I created a separate publication and a different name for this. the goal is to get something interesting. about models Mk4 - super kawaii diffusion Mk3 - additional style prompts - YTM, yutamix-twinkle Mk2 - very strong common style anime checkpoint additional style prompt - YTMmk1 has very cute style, but not good at fingers stuff, and sometimes may be overbright, for example nilou pictures with overbright tags <code>Score_9, score_8_up, score_7_up, masterpiece, best quality,</code> <code>Hires х2 used - r esrgan4xplus-anime6b 10 steps, 0.35 denoise, vae sdxl fp16 fix vae.</code><code>formula 510X990 X 2 hires fix on 10 steps = 1080x1920 this formula maximum effective for use saving both quality and speed. i recommend this formula on all sd models, not just this one.</code><code>i recommend use models only this settings - euler a, hires r esrgan4xplus-anime6b</code>a known bug on forge or other Ui's that you may encounter very rarely or maybe not - you can get an image out of noise, <a target="_blank" rel="ugc" href="https://github.com/hako-mikan/sd-webui-prevent-artifact">try this fix from webui</a> or try change prompts, remove latest prompt what triggered it. mostly it can be triggered by some nsfw prompts or long weird promts with (bla bla bla bla bla bla:1.2) for example

iselestia Diffusion (5 different styles) PXL

iselestiaDiffusion5_mk2PureAnime.safetensors

iselestia Diffusion (4 different styles) PXL

Files with SHA256: 9cee76b223cc45716651c919bd7a4e714744aaad20498e97c726da07e52cb51b