language:
- en
tags:
- stable-diffusion
- text-to-image
license: creativeml-openrail-m
inference: true
High saturation is much less frequently an issue in this version at 7 CFG than the last. But if it happens again, it helps to pull back to 6.
<neg-sketch-2> negative embedding highly recommended for realism and 3D style images (among others). It can be found here: https://huggingface.co/JPPhoto/neg-sketch-2
When prompting for paintings, I suggest using "framed, borders, photo" as your negative prompt to get fullscreen images and cut out any weird 3D-like people.
When prompting for illustrations, I like to use "photo" or "realistic" as my negative prompt.
When prompting for realism, I normally use a negative prompt of <neg-sketch-2> at 1.1 weight and "(anime, render, pixar, illustration, sketch)" at 1.2 weight.
Garbage-bin concepts LoRA recommended for any intense silliness.
Example images hosted on Civitai were generated in InvokeAI's Nodes using latent upscaling from close to ~512x resolution up a few hundred pixels each side at 0.55-0.65 strength w/ DDIM. This was followed up with an ESRGAN model upscale, then converting the image to latents and using ControlNet Tile in a latent to latent stage at 0.2-0.4 strength w/ DDIM.