FRANKENWEIGHTS
dunh dunh dunh...
You wouldn't think something like this might work, but it does. I took the text encoder from my "Storytime" model
and then blindly pasted it over the text encoder stuff in the SD 1.5 model and now we have
It's ALIVE!!!:
Also included here is a somewhat "refined" version of the "Superposition" ComfyUI workflow called This includes
the extra awesomesauce done by ByteDance with their work on
and their
for SD1.5 models, which allows you to greatly reduce the steps required to produce an image --
Frakenweights in this context responds best to "normal" CFG ranges (5-8), especially with the Lora linked above. I generally like to make things that will take a huge range of CFG values, but in this case everything hums along together so well for my purposes that I'm not gonna mess with a good thing. We hack stuff and just try things in these parts, and it just so happens that the combo of FrankenWeights, that Lora, and the "Superposition" bits make a truly Frank-en Stein with enough ummph to bring all your wild ideas to life:
(all of the above come from the same prompts and settings)
So, get your grubby little hands (I'm sure space aliens consider us to be "grubby" and "little") on FrankenWeights today. I have intentionally left this as a 32 bit model, because more often than not the "burn" in an image can be stretched like tasty taffy back into reasonable colors using various image editing things like Photoshop. 32 bits per pixel gives a huge range of latitude for adjusting colors in such programs. I imagine most colorists would be in heaven if there was some video codec that could work at 32 bits. So, welcome to, uh, "heaven" and go to town on even your burnt images because there's all sorts of latitude with 32 bits.