how to avoid Double heads in images

Hi,
I know it is frustrating to get a perfectly generated image but with TWO heads or particularly HEAD over HEAD.
crazy, huh?
now first we have to understand that WHY it happens.

this is an issue with SD1.5 mostly and maybe other models much less (they have other type of issues)

the SD1.5 model BASE is 768x512 and around that.
so the SD1.5 MODELS are mostly trained on and around that but we also have the options to go
CUSTOM 1024 X 1024.
which is a leg above the capability of SD1.5

SD1.5 REFINED models are STILL the best for human anatomy related images.
since the SD1.5 training resource are QUALITY data that is FREE in all DOMAINS. (get it?)
all other models after SD1.5 have CURATED datasets. that means they are Nerfed in creative domains (SFW).

& the BEST way to generated HIGH resolution images using SD1.5 models is to use the UPSCALERS.

so what is with the DOUBLE HEAD? and how to avoid it?

simple.
Ai is expecting the "CANVAS" or area to work around 768 pixels.
anything above it and it assumes that it is ANOTHER connected CANVAS.
so if you make your image size as 1024 tall or 1024 wide, it will do two things in most cases.
on the tall images, it will cause DOUBLE heads.
and on the wide size images, it will create TWINS.

how to avoid that?
Quite a few ways.

A. change the size to 768 instead of 1024 and use UPSCALER to compensate for the size.
you can get 2048 or 2k or 3k or 4k? images as PRO and 1536 X 1024 as Non-PRO on tensor.

B. use uncommon size such as 1010 or 900 or 960 and such.
anything that is avoiding 1024.

C. use different SEED. or steps or CFG.
Technically, Ai art is just MATH equations.
so changing the seed will change the results.
remember,
using the same settings will ALWAYS give u same results!
Ai art is not subjective.
Ai art is purely mathematical.
it is just that we can see that it looks like a woman or a cat.
but Ai is only seeing the CODE. (matrix, huh)

D. use CONTROLNET.
Openpose, image to image, canny, lineart, scribble etc.
controlnet will force the new generations to follow the images u give as input.
you can control how much your new generation should follow the SAMPLE provided.

give it a try.