Type | |
Stats | 796 |
Reviews | (113) |
Published | Mar 23, 2023 |
Base Model | |
Hash | AutoV2 DDEE26637E |
This is my first attempt to create a model. I called it BeautyRaccoonMix.
Main purpose of this model
Originally, the intent of this model was to generate images of cute girls in a style that I found really cool (you can see examples of girls in this style in the sample pictures). BUT! The resulting model exceeded all my expectations! It can generate a variety of things in the same sharp semi-realistic/rendered style, and I think this is awesome! Personally, I had a lot of fun generating my old cool prompts using this model!
Sometimes this model gens such a beautiful eyes!
This version can definitely generate NSFW content, but it's mostly soft. For hardcore/sex-related content, you will need other instruments such as Lora and Hypernets to make this work. I'm planning to create a hard NSFW version, so stay tuned!
How to use
VAE - Be aware, for now v1 is WITHOUT VAE. So if you want to achieve the same results as me, then you might want to use 'vae-ft-mse-840000-ema-pruned'. VAE from 'Any' or 'Waifu Diffusion' also works, but I prefer 'mse-840000'. A version with baked VAE will be coming soon!
PROMPT - Well, you can use any prompts you like. It works with sd and danbooru tags.
NEGATIVE PROMPT - Same here, you can use any negative prompts you like. Personaly I use this one - lowres, bad anatomy, bad hands, text, error, missing fingers, bad fingers, extra fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, bad face, bad eyes. Also embeddings such as easynegative and bad_prompt_version2 works well too. But i think bad_prompt_version2 works a little bit better.
CLIP SKIP - I use clip skip 2. But it seems like 1 and 3 works well too, so you definitely could experiment with it!
CFG SCALE - I use 7. But you can use any values from 5 to 20.
SAMPLER - Mostly I use Euler a with 28 steps. Not really tested on other samplers but it seems ok.
IMAGE RESOLUTION - I prefer gen images on lower resolution (such as 384x512 or 512x384) with Hires. fix and 4x_foolhardy_Remacri upscaler with 0.3 strength.
P.S.
I hope you will enjoy this model same as I did. Share your thoughts and works in comments!
Follow me on twitter to get all recent updates - artraccoonee