Home Stable Diffusion: Part 2
Post
Cancel

Stable Diffusion: Part 2

Welcome to a second part for Stable Diffusion where I give some advices I found or learnt to make better generated images.

Default prompt

For the prompt, I always start with what I give below, it’s a base that works very well for me.

Positive

(For a realistic image) ((realistic, photorealistic, raw photo)), ((masterpiece)), (8k, high quality, ultra high res), (highly detailed:1.2)

(For a painting image) ((realistic painting style)), ((masterpiece)), (8k, high quality, ultra high res), (highly detailed:1.2)

Negative

low res, ((bad anatomy)), (low quality, worst quality:1.4), (monochrome:1.1), (lowres:1.1), disfigured, poorly drawn face, mutation, mutated, (low quality, worst quality:1.4, multiple limbs, creepy face, weird face, more than 2 hand, more than 2 arms, more than 2 legs, more than 2 feet, merged limbs, weird face, distorded body, distorded face), (missing finger, extra digits, fewer digits), ((mutated hands and fingers)), text, signature, deformed, ugly, poorly drawn hands, blur, blurry, (bad eyes:1.2), (misfigured pupils:1.2)

Models

I tested a lot of models to find the one I like the most for different image types.

My most used models so far

Models that I like and use sometimes

Negative embeddings ?!

I know about the negative embeddings, but I don’t use them that often but I still try when I want to see if it can fixes an image I like.

Making upscaled version of an image with more details

One thing I found out is that images upscaled using 4x-UltraSharp are still bad if you zoom in a lot but the problem is that I can’t make draw image more than 512x512 pixels without having an out of memory error.

For that, I installed ControlNet with its models and a script named Ultimate SD Upscale. I use control net to keep the img2img near the original with lowered denoising (0.15 seems like a good value but I check everytimes with X/Y/Z plot script to see the best value) and ultimate sd upscale will cut down the image into 512x512 chunks and upscale with the help of 4x-UltraSharp and add details to the final image.

Sources

This post is licensed under CC BY 4.0 by the author.
Contents