General Comments:

This study presents a novel conditional flow-matching method for SST gap-filling. The authors provide a clear exposition of the method, extensive evaluations, ablation studies for various design choices, and the comparison against prior state-of-the-art ML approaches suggests this method has strong potential for operational applications. This study is of high quality and novelty and the extensive evaluations and ablations make this a particularly strong contribution. With some minor revisions (see below) it is suitable for publication in GMD. 

Congratulations on a great paper and I look forward to seeing it published in GMD.

Best wishes,

Scott A. Martin

Specific Comments:

1. The proposed method re-injects the raw observations back in pixels where they are available. While this ensures small error against available observations it is not necessarily desirable. Observations contain sensor noise and errors themselves, which under this approach will propagate into the reconstructed SST fields. Secondly, this approach will induce strong artifacts in each ensemble member reconstruction at the border between observed and unobserved regions, complicating downstream scientific analysis and interpretation. I think this limitation is just inherent to the approach chosen in this study, so I'm not suggesting the authors modify the method, but the manuscript would benefit from a discussion of this limitation and the motivation for this design choice somewhere in the text.
2. The SST reconstruction methods benchmarked against here (as in the CRITER paper) are all experimental ML methods. While this helps to establish DIRECT as state of the art among ML approaches, the reader is left unclear on whether any of these methods outperform non-ML approaches like objective analysis used in real operational L4 products. The manuscript would be significantly strengthened by including one or more of the operational L4 products (e.g. from CMEMS or PO.DAAC) in the evaluation to the extent possible.
3. DIRECT leverages only the sparse, high-resolution L3 SST observations but microwave sensors provide additional observations that penetrate clouds but with coarser spatial resolution. It would be interesting to consider how the DIRECT approach could be modified in future to incorporate these valuable mesoscale-resolving constraints beneath clouds. Not suggesting the authors modify the method for this study, but it would be an interesting future research direction and could be discussed in the Discussion section.
4. Figs 10 & 11 present isotropic PSD of SST. My understanding is these were computed by taking 2D spectrum of total SST field and then azimuthally averaging? There appears to be a large-scale anisotropic SST background in the ROI which will distort the inferred isotropic PSD. I think this may be part of why all the curves lie so closely on top of each other at all but the highest wavenumbers. A more appropriate metric would be to compute these spectra on the SST anomaly from a coarse background (e.g. linear plane fit to the ROI?).
5. Figs 12 & 13: the pixel scales are hard to interpret physically. Either replace with km or add an additional approximate km reference scale to help orient the reader.
6. The CRPS values in Table 3 are not particularly informative without any form of reference to compare against. I'm not familiar with CRITER, but you present estimated uncertainty from that method so why can't we also get a CRPS for it? I'd honestly suggest just removing CRPS from the paper if there is nothing to benchmark against, another option would be to include MAE of the deterministic methods as reference values in the table since CRPS reduces to MAE for single member predictions?
7. Multi-panel figures like Fig 6 throughout the paper would benefit from clearer sub-panel labels (e.g. panel a, b, etc.) and panel 6 is particularly hard to follow at first glance.
8. Line ~109: typo "networ".
9. The fact the method can be trained exclusively on sparse real observations is to my mind a major strength compared to existing literature and could be better emphasized with a prominent mention somewhere in the abstract.

This manuscript addresses an important problem and proposes a reliable method for reconstructing SST fields with uncertainty estimates. The experiments are extensive and generally convincing. I recommend a minor revision, with the following comments.

1. The Introduction should include recent work on spatiotemporal diffusion models for climate and SST reconstruction, for example, Qian et al. (2026, arXiv:2602.16515) and Li et al. (2026, arXiv:2603.16272). A brief comparison would help clarify DIRECT's contribution relative to these studies.

2. Figure 2 and the corresponding description in Section 2 are somewhat confusing and difficult to follow. In particular, (x{t-1}) and (x{t+1}) are pixelwise composites from several dates, while (M_{t-1}), (M_t), and (M_{t+1}) have different meanings. A clearer schematic, example, or brief pseudocode would be helpful. The fixed three-day search range and nearest-valid-observation rule are also rather empirical, so their motivation and possible limitations should be discussed more explicitly.

3. Cloud masks from other dates are used to construct training samples. This is practical, but cloud occurrence may be coupled with atmospheric conditions and SST variability. Therefore, the missingness mechanism is not necessarily independent of the target SST field. The authors should discuss this assumption and clarify that the evaluation is based on synthetically withheld observed pixels rather than direct validation beneath real clouds.

4. The variance correction provides a lightweight global calibration, but it does not necessarily demonstrate good pixelwise or conditional calibration. I suggest reporting the empirical coverage of nominal, for example,  50%, 80%, 90%, and 95% prediction intervals on the test set. Coverage before and after the correction would be particularly informative.

5. The satellite SST product is treated as ground truth, although it has its own retrieval uncertainty. I understand that explicitly incorporating observation errors would require substantial methodological changes, but this issue should at least be discussed because the reported residuals may include both reconstruction and observation errors.

I strongly recommend adding a short Discussion section covering the empirical temporal-context construction, the use of transferred cloud masks, the global nature of the variance correction, and the neglect of observation uncertainty.