Closed Enferlain closed 2 months ago
This method was something I stumbled upon by chance during various trial and error processes. Therefore, there is no theoretical explanation for it. It involves tinkering with special elements at the entrance and exit of U-Net. Specifically, the following elements will be modified:
model.diffusion_model.input_blocks.0.0.weight"
model.diffusion_model.input_blocks.0.0.bias
model.diffusion_model.out.0.weight
model.diffusion_model.out.0.bias
model.diffusion_model.out.2.weight
model.diffusion_model.out.2.bias
Are these adjusting some individual layers known to be responsible for the described result? Like detail/noise etc? How did you find out which layers they are, and does it apply to sdxl?
I can't find the names of the layers in the code :(