[1] Stage B is the most complex and expensive part of the method, but it adds very little to the final result. When using the groundtruth (oracle), Stage A accounts for about ~95+% of the total PSNR gain while stage B adds only +0.04 dB PSNR and −0.005 LPIPS on top of Stage A. The deployable Pred-H model is also almost the same as the Stage A oracle row (+0.009 dB PSNR, −0.001 LPIPS). The effectiveness of the stage B is highly unclear.
[2] The projection step is the core idea of Stage A, but no ablation is provided on that. At least some experiments should be provided in which stage A is trained with the plain update instead of projected one. Since most of the gain comes from Stage A (see [1]), it is unclear whether the gain comes from the projection or simply from additional training on Base outputs. Without this ablation, contribution 2 is not supported.
[3] The role of the evidence map is unclear. Pred-H performs almost identical to Oracle-H. This could mean the predictor is very accurate, or it could mean the model largely ignores the map. I reckon the second case is quite possible, given how small the Stage B gains are. The authors should report results with no map, a constant map and a random map. They should also report the accuracy of the evidence predictor.
[4] The obvious simple baselines are missing. For example: running the Base a second time, fine-tuning the Base for longer, or training a Base-conditioned refiner with a standard loss. Figure 1 claims that repeated restoration fails, but this is never measured. There is also no comparison against plain RL fine-tuning of the same Base. Without these, it is not possible to conclude on whether the proposed machinery is needed/useful.
[5] The presentation of the oracle results is misleading. The oracle map is computed from the test ground-truth image, so the oracle rows use target information at test time. Even so, Table 1 bolds the Oracle-H values as the best results "across all rows." This makes the method look stronger inadequately and should be avoided for authentic presentation.
[6] The training reward and the evaluation diagnostics are nearly the same. Training rewards repair in high-error regions and preservation in low-error regions, using pixel and LPIPS errors. Table 2 then scores the model with the same kinds of quantities, using region weights from the same oracle map. So the gains in Table 2 are expected by construction. Independent evaluation is needed, for example a user study, regions based on face parsing, or other perceptual metrics.
[7] The qualitative results are very limited and underwhelming. Only two examples in main paper are not enough, the figure does not point out the regions where one could notice improvements if any, and the outputs of the baselines seem to have much lower brightness/contrast compared to that of the proposed method and groundtruth – potentially a bug in generating / displaying the outputs?
[8] The improvements are small and come without any statistical support. In Table 2, Repair Success rises by 1.7 points, Pix-Fail falls by 1.4 points, and patch degradation falls by about 0.01. There are no multiple seeds or confidence intervals.
[9] The experimental scope is limited: Only one benchmark is used (CelebA-Test-Ref with synthetic degradations). There is no evaluation on real-world degraded images. Also, the refiner is applied only to the authors' own Base. The paper frames the problem as a general post-restoration setting, but it never refines outputs from other competing methods.