Hi,
I want to finetune the diffusion model on my dataset. My images are greyscaled, so they have only one channel.
I already successfully finetune it using np.repeat for each image.
But I am wondering is there a way to make it better, faster, and more efficient?
I want to finetune the model without channel duplication (np.repeat).
I tried to reduce weights dimensions using slicing/mean/max. For example, from [3, 128, 128] to [1, 128, 128], from [3] to [1] for bias.
But in this case, after each denoise step, I got the same noize image as it was after adding the noise step. So it did not denoised properly.
I much appreciate any help or suggestion with this. Is the problem in the noise/denoise step or in the weights dimension reducing itself?