All my models are free but if you want to support me you can buy me a coffe :3
Hi, this is a work in progress of the project. I'm sharing it because I have achieved acceptable results.
I know there are other models that do the same, but for me, it has been difficult to obtain realistic results. That's the purpose of this.
In V2 the hands are better, and if you use control-net for hands and some negative emb, i think it give you an acceptable result.
RECOMMENDATIONS
Expect some merged bodies.
Sometimes work if use both tags at same time.
When your composition is good i recommend use hire.fix at 0.3 and restore faces for upscale.
CLIP SKIP: 2
You can install this Extension for automatic1111, for block the weights, put "MIDD" in the lora like this: <lora:Headlessv2:1:MIDD>
(it's the same that put: <lora:Headlessv2:1:1,0,0,0,1,1,1,1,1,1,1,1,0,0,0,0,0> and you can change this numbers(the weights) to get better results)
Here's a comparison of MIDD, V2 and V1:
use at your convenience.
recommendations for only in v1
Dont use tags at same timeFor horizontal or vertical resolutions i recommend reduce it a little : if 512x768 put 512x752
Description
after a lot of tries i just do something unprofessional but it worked so here is.
maybe i upload another fix but not soon.
this version has:
Better hands (not always but i think you can fix it easily now
Better prompt adaptation (in my test i generated 9 images per prompt and at least 2 got the concept)
Better model adaptation (see the images)
Not always worked, and the diembodied head its a little tricky