Question: DC layer replaces k-space samples with k0 — assumption in SR setting?

[030108ming]

Hi

I'm implementing DiffMSR and have a question about the data-consistency (DC) layer.

In the code, the network prediction pred (complex image, 2ch real/imag) is FFT'ed to k-space (k), then DC enforces measured samples:
out = (1 - mask) * k + mask * k0 (noiseless)

This suggests that at inference time we have access to:
(1) k0: measured/known GT k-space samples
(2) mask: sampling pattern / locations of those samples

My confusion is about the super-resolution / low-resolution setting:

• In many practical SR cases we only have LR reconstructed images, not the GT k-space samples.
• In your experiments (e.g., k-space 64×64), is k0 obtained by cropping/subsampling the full k-space (synthetic LR), or do you assume raw acquired k-space is available?
• If only LR images are available (no k-space), how should DC be applied in practice? Skip DC, approximate k0 by FFT(LR), or is the method intended only when k-space is accessible?

Any clarification would be greatly appreciated. Thanks!