🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑
First of all I would like to start off by saying this is a experiment in tricking Stable Diffusion. I have longed for mixed race bodies that more resemble what I am fond of, bodies like my significant other. SD didn't seem to do it right, and was had to control to get the effect I wanted.
I also didn't want to release a model that would reproduce identifiable faces for privacy and respect, and just wanted the artistic form of the generalized body type.
To defeat this I put together an idea to experiment with:
Gather images that represent the dataset I wanted
Specifically crop images so that only the body is in the frame.
Upscale the images with 4x-UltraSharp.pth, downsize to 512x512
3.a For images that the face could not be cropped out, I masked and filled with latent noise (using ComfyUI and my WAS Node Suite)Train with prompts that focus on the body type I am going for
4.a Add specific emphasis on faces, which don't exist, again in the style I am going for
The idea here is that while training, and using my prompts, Stable Diffusion is going to be force to make up unique faces on it's own. To me the frequency of cut-off faces seems very low, and so far, my theory seems to have panned out.
🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑🤎🍑
Notes:
Best used with VAE vae-ft-mse-840000
Best used with good ol' negative embeds for composition, and bad hands
Best used with models like Dreamlike, Protogen, TheAlly's Mix's, etc.
Due to the nature of the TI, if you want clothed models, you'll have to play with weighting to taste.
Description
FAQ
Comments (8)
so Comfy is still alive and kicking?
Yeah. It has issues, like default InvokeAI style weighting which is garbage, so need the adv clip text encode. I put up a plugin list node here: https://github.com/WASasquatch/comfyui-plugins/blob/main/README.md
@WAS cool 10x i'll check it out
I keep having issues with different versions, by the time I fixed some missing nodes, now I'm getting
'Failed to validate prompt for output 189 Parsing error's
I'll try go over these
@iclamalo The failed to validate is usually a input missing on a node somewhere. And yeah, if a node is renamed, it just breaks workflows. Kinda annoying. I wish Comfy used the class names to identify nodes and whatever the title was, could be whatever. Now he added "safe names" which is just a whole new list to keep track of duplicating most of what you have in the node mapping list for titles. Some backwards logic here and there. But still powerful.
@WAS he's not so convenient, mr comfy
You're saying that the TI creates a mixed heritage or darker complexioned body, and that "forces" SD to create a matching head to attach? Diabolical!
Oh no, I included extensive descriptions of the face/head even though they were cropped out, so when it's sampling with prompt, SD is getting a lot of prompting regarding a head that doesn't exist, so it makes it kinda rearrange things and do a head. During sampling maybe a 5-10% came out cut off, which can show in renders, but for the most part, they all had faces during sampling and in generations. Some models don't do too well though. Redshift had the most cut-off heads and bodies halfway in the frame, but not sure if that's because of the original 1.4/1.5 issue with cut-offs.
@WAS Ah, I see. It's still pretty clever.