Original one has sparse inverse operation to achieve 2nd equation of formula (12) in original paper, while yours doesn't. Is there any change applied on formulas?
Your implementation has multiple stages of U-Net which doesn't appear in original paper. Instead, I found this paper has similar usage of U-Net as denoising network here. Did you use U-Net as some kind of equivalent to part of the original ADMM-net or sth else?
I found some differences in the original ADMM-net and your implementation