to_rgb branch has only 1 learnable kernel

[inspirit]

to_rgb Conv has only 1 learnable kernel but we add more here:
https://github.com/lucidrains/gigagan-pytorch/blob/9a364dd33cf0ff053ea01041e02bc41949e53609/gigagan_pytorch/gigagan_pytorch.py#L995C20-L995C20

[lucidrains]

gigagan-pytorch/gigagan_pytorch/gigagan_pytorch.py

Lines 896 to 897 in 9a364dd

	is_adaptive = num_conv_kernels > 1
	dim_kernel_mod = num_conv_kernels if is_adaptive else 0

yup, so that's actually set to 0 here

i may move all that modulations into the adaptive conv2d mod at some point

its too confusing the way i have it currently

[inspirit]

this doesnt fix the issue i'm talking about:
to_rgb = AdaptiveConv2DMod(dim_out, channels, 1, num_conv_kernels = 1, demod = False)
num_conv_kernels is always 1 for rgb regardless of Generator setting

[lucidrains]

@inspirit yup, that is intentional, based on point 2 in #33

[inspirit]

but you still compute kernel weights for rgb branch

style_embed_split_dims.extend([
dim_in, # for first conv in resnet block
dim_kernel_mod, # first conv kernel selection
dim_out, # second conv in resnet block
dim_kernel_mod, # second conv kernel selection
dim_out, # to RGB conv
dim_kernel_mod, # RGB conv kernel selection
])

[lucidrains]

@inspirit 🤦‍♂️ yes you are right

[inspirit]

but i agree that this path looks super convoluted :) i would rather move mod+kernel_mod projection inside AdaptiveConv2DMod

[lucidrains]

@inspirit yes i'll get around to it 😅