Self-Supervised Selective-Guided Diffusion Model for Old-Photo Face Restoration

Thank you for recognizing the novelty and rationality of our SSDiff framework, including the idea of pseudo-labeled references, the clear writing, and the effectiveness of the selective guidance strategy. Below, we address your concerns point by point. As the current rebuttal system does not provide a way to include figures, some clarifications are given quantitatively, and the corresponding visualizations will be added in the final version.

• Weak1: Quantitative evidence. Yes, our method is applicable even when old photos have no visible breakage. When masks are empty, these modules act as normal restoration, and the later-stage color refinement still benefits faded images. Table S1 in our supplementary reports results on WebPhoto-Test (407 faces) and CelebA-Child (180 faces), both consisting of old face photos without breakage. SSDiff achieves the best or second-best scores across quantitative results, including FID and NIQE on the above benchmarks, confirming its effectiveness beyond broken cases with quantitative evidence.

• Weak2: Robustness of face-parsing maps. The performance of SSDiff is generally robust to inaccuracies in the face parsing network. In our framework, the pre-trained parsing model is used only to provide a coarse directional signal during the reverse diffusion process, similar to the loose guidance mechanism in classifier-guided diffusion, rather than as a strict constraint. As long as this signal is not severely misleading (i.e., without gross parsing failures), the strong generative prior of the frozen diffusion model dominates the reconstruction and can recover plausible facial structures.

  As discussed in Sec. C.1 of the supplementary, face parsing is considerably more robust than facial landmarks or 3D face models. Based on extensive experiments, we have observed that parsing maps almost never fail completely, even for degraded faces, and typically provide a sufficiently accurate spatial prior. To quantify this, we manually degraded 100 randomly selected high-quality CelebA‑HQ faces using the degradation model in GFPGAN [1], where degradation parameters (blur kernel σ, downsample scale r, Gaussian noise δ, JPEG quality factor q) are randomly sampled from (0.2:10, 12:16, 0:15, 60:100). These degraded images can be classified as hard-degraded following the standard of Fig. S1 in the supplementary, still achieve an average face contours IoU of 79% between their parsing-map and those of original high-quality images.

  Qualitatively, across the visual results in the main paper, some parsing maps are clearly imperfect, yet the restored faces remain visually stable and of high quality. For example, in Fig. 6 (second row), part of the face contour is occluded, leading to an inaccurate parsing map, but the final restoration remains natural, demonstrating that SSDiff does not rely on precise parsing boundaries.

  To further evaluate robustness against parsing-map inaccuracies, we conducted a controlled experiment on the Medium subset of VintageFace by replacing the original parsing maps with pseudo-label maps generated with different guidance strengths $s$. The resulting face contours IoU with respect to the original parsing map are:

  • for s=2.5e-4, IoU ≈ 76% (24% discrepancy);
  • for s=1e-4, IoU ≈ 70% (30% discrepancy);

  The resulting performance is summarized below:

  Parsing Maps	Ours	IOU(76%)	IOU(70%)
  FID($\downarrow$)	128.3	131.1	133.4
  MAN-IQA($\uparrow$)	0.395	0.391	0.382
  Face Similarity ($\uparrow$)	1	0.985	0.955

  Here, Face Similarity (range [0, 1]) denotes the cosine similarity between features (extracted with ArcFace [3]) of the perturbed restoration and the original restoration (Ours). This result shows that errors in the parsing map network are not severe (Less than 40%), and the strong generative prior of the diffusion can propagate the correct cues to other face regions, preventing significant performance drops. This demonstrates that SSDiff is robust to these pre-trained networks. We will include these quantitative results and additional visual examples in the final version.

• Weak3: $T_1$ range. Thank you for this suggestion. In extensive experiments on our whole benchmark, we also find that the transition point $T_1$ is not a fixed value but lies within a range. However, for simplicity and consistency across all experiments, we set $T_1=0.4T$, which corresponds well to most cases. In practice, depending on the degree of degradation, $T_1$ typically falls between $0.35T$ and $0.45T$; lighter degradations tend to shift $T_1$ slightly bigger. We will clarify this in the final version and include additional visual examples to illustrate the effect of this variation.

• Weak4: Selection of non-face breakage regions. For selecting breakage regions outside facial components, our method combines the scratch mask with the labels from the face parsing map. The parsing map identifies non-face areas such as background, skin, and hair, while the scratch mask (binary: white=1 for breakage, black=0 for intact) localizes breakage regions. Taking the intersection of these two masks gives the breakage regions that do not contain facial components, which are then guided for restoration. We will incorporate these instructions into our final version.

• Weak5: Why CLIP-based face identity distances? Traditional identity distances metrics rely on facial landmark detection (e.g., ArcFace [2]) and domain-specific face recognition embeddings. In input old photos with breakages, landmarks are often missing or heavily distorted, which leads to unstable or even invalid identity features and thus unreliable distances. To avoid this bias, we adopt CLIP-based face identity distances. CLIP embeddings are computed directly from the entire image without requiring landmark alignment, and they are trained on very diverse, cross-domain image-text data. This makes them more robust to blur, fading, and partial breakage, allowing us to measure identity similarity more stably and fairly under the challenging conditions of old-photo restoration. We will incorporate the above discussion into our final version.

• Weak6: Details error. Thank you for your careful review. We will correct $s$ = 3.5e-3 to $s_s$ = 3.5e-3 in the implementation details of our main paper.

[1] Wang X, Li Y, Zhang H, et al. Towards real-world blind face restoration with generative facial prior. CVPR 2021.

[2] Deng J, Guo J, Xue N, et al. Arcface: Additive angular margin loss for deep face recognition. CVPR 2021.

Thank you for highlighting the insightful analysis of guided intensities and the novel idea of using weakly supervised generated images as references for color and texture. We address your concerns point by point below. As the current rebuttal system does not provide a way to include figures, some clarifications are given quantitatively, and corresponding visualizations will be added in the final version.

• Q1: Calculation of FID. The calculation of FID in our experiments follows official codes of VQFR [1] rather than computing statistics only on 300 images in our VintageFace. Specifically, we use the pre-trained inception weight released with VQFR, which is trained on 70k high-quality FFHQ images. FID is then computed between the feature distribution of our restored images and that of the reference FFHQ distribution. Therefore, the reliability of FID does not depend on the number of images in VintageFace itself, but on the fixed statistics from the large-scale FFHQ reference.

• Q2: Reference metrics on synthetic data. Due to time constraints, we randomly select 1000 face images from CelebA-HQ for evaluation. Synthetic degradations are applied using the same degradation model as GFPGAN [2], where degradation parameters (blur kernel σ, downsample scale r, Gaussian noise δ, JPEG quality factor q) are randomly sampled from (0.2:10, 1:8, 0:15, 60:100) to form a mixed-degradation set. We report LPIPS and DISTS to assess perceptual quality and include face landmark distance (LMD) as a fidelity metric (as in VQFR [1]). PSNR and SSIM are not used as they primarily measure pixel similarity rather than face identity preservation.

  CelebA-HQ (synthetic)	DifFace	DiffBIR	DDNM	PGDiff	Ours
  LPIPS ($\downarrow$)	0.2722	0.2659	0.4012	0.3104	0.2695
  DISTS ($\downarrow$)	0.1441	0.1431	0.3281	0.1636	0.1320
  LMD ($\downarrow$)	3.37	3.01	3.52	4.15	3.29

  Compared with existing diffusion-based methods, our method achieves competitive performance, only behind DiffBIR, which is trained on large paired datasets generated with the same degradation model. Note that our focus is on old-photo face restoration rather than blind face restoration. Nevertheless, as shown in the above tables and in our paper, SSDiff shows strong robustness to synthetic degradations. These results indicate that SSDiff is capable of handling both old face photos and synthetic degraded faces, whereas methods such as DiffBIR, though effective for synthetic degradations, struggle with old face photos.

• Q3_1: Robustness of pre-trained networks. SSDiff is generally robust to inaccuracies in external components such as face parsing and scratch detection networks. These pre-trained networks provide only coarse directional guidance during reverse diffusion—similar to classifier-guided diffusion—rather than strict constraints. As long as their outputs are not severely erroneous, the strong generative prior of the frozen diffusion model dominates reconstruction and recovers plausible structures. Consequently, even when these auxiliary outputs are imperfect, the restorations remain stable and of high quality. Below, we analyze visual cases present in our paper to prove this point:

  Parsing maps: As shown in Fig. 6 (second row) of the main paper, even with occluded contours, leading to inaccurate parsing, the results remain visually natural. Furthermore, as discussed in our analysis of "Weak2" for Reviewer pXcG, even under severe degradations, the IoU between degraded facial contours and GT contours remains around 79%, demonstrating the robustness of the parsing network.

  Scratch detection: As shown in Fig. S4 (second row) of our supplementary, large stains on the head are only partially detected, yet the restoration remains natural. Similarly, in the failure case in Fig. 11, only a fraction of the stains are detected, but the restored results are still better than other methods. This demonstrates that the model uses detected regions as a rough guide and can roughly recover nearby features even with incomplete masks.

  To quantify this robustness, we conduct additional experiments on the Medium degradation subset:

  1. Parsing maps: We introduce inaccuracies by replacing the original parsing map with the pseudo-label parsing map of different strengths $s$, where the resulting errors are even larger than those observed in parsing maps under severe degradations (79% IOU). The resulting facial contours IoU with the original parsing maps is:

     • for s=2.5e-4, IoU ≈ 76% (24% discrepancy);
     • for s=1e-4, IoU ≈ 70% (30% discrepancy);

     Parsing Maps	Ours	IOU(76%)	IOU(70%)
     FID($\downarrow$)	128.3	131.1	133.4
     MAN-IQA($\uparrow$)	0.395	0.391	0.382
     Face Similarity ($\uparrow$)	1	0.985	0.955

  2. Scratch masks: Randomly flip 10%, 20%, and 30% of the masks of breakage regions to simulate a situation where some of the breakages have not been detected.

     Scratch Masks	Ours	10% flip	20% flip	30% flip
     FID($\downarrow$)	128.3	129.2	130.7	132.5
     MAN-IQA($\uparrow$)	0.395	0.391	0.379	0.381
     Face Similarity ($\uparrow$)	1	0.974	0.953	0.937

     Here, Face Similarity (range [0, 1]) denotes the cosine similarity between features (extracted with ArcFace [3]) of the perturbed restoration and the original restoration (Ours). This result shows that errors in the parsing map network are not severe (Less than 40%), and the strong generative prior of the diffusion can propagate the correct cues to other face regions, preventing significant performance drops. This demonstrates that SSDiff is robust to these pre-trained networks. We will include these quantitative results and additional visual examples in the final version.

• Q3_2: Computational Costs & Speed. As detailed in Sec. C.2 and Sec. C.3 of our supplementary, we have analyzed the impact of the auxiliary components (e.g., pre-trained face parsing and scratch detection networks) on both computational cost and inference time:

  Params and FLOPs: As shown in Table S2 of our supplementary, our SSDiff requires fewer parameters and FLOPs than existing diffusion-based methods. Because the face parsing and scratch detection networks can be executed offline, reported values in the supplementary do not include them. For completeness, if we include these components, the total becomes 105.4 M parameters and 140.6 G FLOPs, still significantly lower than DifFace (175.4 M, 268.8 G) and DiffBIR (1717 M, 23234 G). PGDiff also depends on scratch masks; using the same scratch detection model, its cost would be roughly 84.3 M parameters and 146.8 G FLOPs, comparable to ours. These results indicate that SSDiff does not suffer from a computational disadvantage.

  Inference Speed: As reported in Sec. C.2 of our supplementary, all auxiliary networks are executed once in reverse diffusion rather than at every sampling step. Even when all networks run online, the added latency on an NVIDIA RTX 4090 is only about 95 ms, which is negligible compared to the tens of seconds required by the diffusion sampling itself.

• Q4: Robustness to pseudo-label inconsistency. We appreciate your concern. This issue was carefully considered in the design of SSDiff. As shown in the statistical analysis in Fig. 2 of the main paper, by controlling the strength of weak supervision, we ensure that the pseudo-labels (generated by a frozen diffusion model) closely align with facial contours, with only small deviations (reflected in the reported IoU values). Moreover, these pseudo-labels are not used as pixel-level constraints; they serve only as coarse guidance for color style and smoothness in damaged regions, so moderate inaccuracies have limited influence on the final restoration.

  To further assess robustness under less accurate pseudo-labels, we conducted an additional experiment on the Medium degradation subset using pseudo-labels generated with a smaller sampling step size (s=1e-4), where the average IoU of facial contours with the original input drops from 87% to 70%. As shown below, compared to the original pseudo-labels (s=1e-3), the resulting image quality and Face Similarity (defined in "Q3_1") degrade slightly. This confirms that SSDiff remains relatively robust even when pseudo-labels are partially less reliable.

  Pseudo-label quality	MAN-IQA($\uparrow$)	FID($\uparrow$)	Face Similarity($\uparrow$)
  IOU:87% (Orginal)	0.395	128.3	1
  IOU:70%	0.374	136.6	0.929

[1] Gu Y, Wang X, Xie L, et al. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. ECCV 2022.

[2] Wang X, Li Y, Zhang H, et al. Towards real-world blind face restoration with generative facial prior. CVPR 2021.

[3] Deng J, Guo J, Xue N, et al. Arcface: Additive angular margin loss for deep face recognition. CVPR 2021.

Thank you for recognizing the novelty of our SSDiff, including the self-supervised design using pseudo-references, the staged coarse-to-fine guidance strategy, and the selective restoration mechanism that preserves identity-sensitive regions. Below, we address your concerns point by point. As the current rebuttal system does not provide a way to include figures, some clarifications are given quantitatively, and the corresponding visualizations will be added in the final version.

• Weak1: Limited Benchmark Scale. The VintageFace benchmark contains 300 images, but this benchmark is not a primary contribution; rather, it was introduced because no such dataset currently exists to provide an initial evaluation set for restoration performance under extreme old-photo degradations, including breakage, fading, and severe blur. We plan to further expand this benchmark in future work. Moreover, as shown in Table S1 in our supplementary, we have also validated SSDiff on broader datasets containing old photos (WebPhoto-Test and CelebA-Child), demonstrating that our method is not limited by the current scale of VintageFace.

• Weak2: Narrow Identity Preservation Test. To address your concern, we evaluate on the subset of CelebA-HQ (1000 images) with synthetic degradations generated following the degradation model in GFPGAN [1], where degradation parameters (blur kernel σ, downsample scale r, Gaussian noise δ, JPEG quality factor q) are randomly sampled from (7:9, 8:16, 20:40, 30:40) to form a severe-degradation set (measured using Fig. S1 in the supplementary). We recompute CLIP-based identity distances on this dataset, and we further adopt face landmark distance (LMD) [2] as the face fidelity metric. PSNR or SSIM are not used as they primarily measure pixel consistency rather than face identity preservation. As shown in the table below, our method remains consistent with the paper's findings: SSDiff keeps competitive among diffusion-based methods, only behind DiffBIR on the LMD metric. Note that DiffBIR is trained on large paired datasets synthesized using the same degradation model, whereas SSDiff is training-free.

  Degradation Type	DifFace	DiffBIR	DDNM	PGDiff	Ours
  Identity Distance ($\downarrow$)	0.171	0.154	0.187	0.202	0.153
  LMD ($\downarrow$)	4.20	3.88	4.85	5.19	4.12

• Q1: Robustness of pre-trained networks. SSDiff is generally robust to inaccuracies in external components (face parsing, scratch detection, style transfer). These networks only provide coarse directional signals during reverse diffusion, similar to classifier-guided diffusion, and are not strict constraints. As long as the guidance is not severely misleading, the strong generative prior of the frozen diffusion model dominates reconstruction. Across many visual results in the paper, there are inevitably cases where these networks are imperfect, yet the restorations remain stable and of high quality.

  Parsing maps: As shown in Fig. 6 (second row) in our main paper, even with occluded contours leading to inaccurate parsing, the results remain visually faithful. Furthermore, as discussed in our analysis of "Weak2" for Reviewer pXcG, even under severe degradations, the IoU between degraded facial contours and GT contours remains around 79%, demonstrating its robustness. In addition, the results of the parsing map do not directly supervise the guidance, but rather in an indirect manner, also leading to better robustness.

  Scratch detection: As shown in Fig. S4 (second row) in our supplementary, large stains on the head are only partially detected, yet the restoration remains natural. Similarly, in the failure case in Fig. 11, only a fraction of the stains are detected, but the restored results are still better than other methods. This demonstrates that the model uses detected regions as a rough guide and can roughly recover nearby structures even with incomplete masks.

  Style transfer: Although the main paper does not include intermediate style-transfer outputs, we observe the intermediate style‑transfer result for Fig. 7 (first row). The result shows that transferred images initially have a yellowish facial tint; however, once used as a coarse color prior, the diffusion process refines and propagates realistic colors, producing a natural complexion.

  To quantify this robustness, we conducted additional experiments on the Medium type subset:

  1. Parsing maps: We introduced inaccuracies by replacing the original parsing maps with pseudo-label parsing maps of different strengths $s$, where the resulting errors are even larger than those observed in parsing maps under severe degradations (79% IOU). The resulting IoU with the original parsing map is:

     • for s=2.5e-4, IoU ≈ 76% (24% discrepancy);
     • for s=1e-4, IoU ≈ 70% (30% discrepancy);

     Parsing Maps	Ours	IOU(76%)	IOU(70%)
     FID($\downarrow$)	128.3	131.1	133.4
     MAN-IQA($\uparrow$)	0.395	0.391	0.382
     Face Similarity ($\uparrow$)	1	0.985	0.955

  2. Scratch masks: Randomly flip 10%, 20%, and 30% of the masks of breakage regions to simulate a situation where some of the breakages have not been detected.

     Scratch Masks	Ours	10% flip	20% flip	30% flip
     FID($\downarrow$)	128.3	129.2	130.7	132.5
     MAN-IQA($\uparrow$)	0.395	0.391	0.379	0.381
     Face Similarity ($\uparrow$)	1	0.974	0.953	0.937

  3. Style transfer: We weaken the style transfer guidance by reducing the style factor α from 0 to 0.1 and 0.2, slightly affecting color and content.

     Style Transfer	Ours (α=0)	Poor-1 (α=0.1)	Poor-2 (α=0.2)
     FID($\downarrow$)	128.3	129.2	128.8
     MAN-IQA($\uparrow$)	0.395	0.396	0.392
     Face Similarity ($\uparrow$)	1	0.977	0.959

  Here, Face Similarity (range [0, 1]) denotes the cosine similarity between features (extracted with ArcFace [3]) of the perturbed restoration and the original restoration (Ours). These results show that errors in the pre-trained networks are not severe (Less than 40%), and the strong generative prior of the diffusion model can propagate the correct cues to other regions, preventing significant performance drops. This demonstrates that SSDiff is robust to these pre-trained networks. We will include these quantitative results and additional visual examples in the final version.

• Q2: How to select 50 LQ samples for measuring fidelity. The 50 LQ images used for fidelity evaluation are chosen from samples with relatively mild degradation, particularly those without breakage or severe blur in key facial components such as eyes, nose, and mouth. Since in the absence of ground truth, identity distances have to be computed with input old photos. For heavily degraded images with severe blur, fading, or breakage, a successful restoration even incorrectly increases the identity distance (where a smaller distance indicates better fidelity) because the restoration corrects degradations that are inadvertently treated as part of the original face identity. Therefore, to mitigate this bias, we select old photos with lighter degradation and better-preserved facial features, providing a more reliable evaluation of identity fidelity.

• Weak3 & Q3: Generalization to other datasets and degradations. As shown in Table S1 and Fig. S7 of our supplementary, we have evaluated SSDiff on popular benchmarks beyond our VintageFace, including WebPhoto-Test (407 LQ faces) and CelebA-Child (180 LQ faces). These sets contain a large number of historical face images with varying degrees of blur and fading but no breakage, where our masks automatically degenerate to standard restoration (e.g., all black). SSDiff achieves superior quantitative (FID, NIQE) and qualitative results on both benchmarks, demonstrating strong generalization.

  Following your valuable suggestion, we further performed additional evaluations on a 1000-image subset of CelebA-HQ using the same degradation model in GFPGAN [1], where degradation parameters (blur kernel σ, downsample scale r, Gaussian noise δ, JPEG quality factor q) are randomly sampled from (0.2:10, 1:8, 0:15, 60:100) to form a mixed-degradation set. We further evaluate on the real benchmark you mentioned (LFW-Test with 1711 LQ faces).

  CelebA-HQ (synthetic)	DifFace	DiffBIR	DDNM	PGDiff	Ours
  LPIPS ($\downarrow$)	0.2722	0.2659	0.4112	0.3104	0.2695
  DISTS ($\downarrow$)	0.1441	0.1431	0.3281	0.1636	0.1320
  FID ($\downarrow$)	69.3	51.2	115.8	58.1	50.3
  NIQE ($\downarrow$)	4.29	6.08	8.83	4.837	4.16

  LFW-Test (real)	DifFace	DiffBIR	DDNM	PGDiff	Ours
  FID ($\downarrow$)	46.4	40.4	112.3	47.3	42.2
  NIQE ($\downarrow$)	4.30	5.74	9.64	4.11	4.03
  MAN-IQA ($\uparrow$)	0.368	0.423	0.275	0.373	0.411

  As shown in these tables, SSDiff consistently achieves the best or second-best results across both reference and non-reference metrics. Note that our focus is on old-photo face restoration rather than real-world face restoration. Nevertheless, as shown in the above tables and in our paper, SSDiff demonstrates strong robustness to real-world degradations. These results indicate that SSDiff is capable of handling both old face photos and real-world degraded faces, whereas methods such as DiffBIR, though effective for real degradations, struggle with old face photos.

[1] Wang X, Li Y, Zhang H, et al. Towards real-world blind face restoration with generative facial prior. CVPR 2021.

[2] Gu Y, Wang X, Xie L, et al. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. ECCV 2022.

[3] Deng J, Guo J, Xue N, et al. Arcface: Additive angular margin loss for deep face recognition. CVPR 2021.

Thank you for recognizing the clear idea of our method, the staged guidance design, and the soundness of our analysis, as well as the strong performance of SSDiff on benchmarks. We also appreciate your suggestions for adding more visual examples. While the rebuttal system does not allow us to include additional figures, we provide quantitative clarifications where possible and will incorporate your recommended visualizations in our final version.

• Weak1 (a): Explanation of Table 2. The baseline without our self-supervised strategy still benefits from two factors: 1. We retain the L1 loss, which provides a coarse directional signal that makes restorations visually clearer. 2. Mask-guided diffusion explicitly marks unreliable breakages so that frozen diffusions can use its generation to hallucinate missing content. Although this can only recover simple breakages, it naturally leads to better results than methods without any breakage repair. However, these results should not be confused with the self-supervised strategy. By introducing pseudo-reference guidance, it substantially improves color, texture, and fine details.

• Weak1 (b): Results in more scenarios. Thank you for this suggestion, which makes our evaluation more comprehensive. Below, we evaluate our SSDiff with existing diffusion-based methods:

  More poses: We collect 20 relatively profile-view faces (fully side profiles are rare) from VintageFace and the web, and 20 frontal faces with comparable degradation levels (measured using Fig. S1 in the supplementary). The metrics show no significant gap between the two settings, demonstrating that such cases do not affect the robustness.

  Poses	Front	Profile-view
  FID($\downarrow$)	128.6	132.9
  MAN-IQA($\uparrow$)	0.3975	0.3920

  Furthermore, we evaluate 100 fully profile-view faces from real benchmarks (LFW-Test, WebPhoto-Test, Wider-Test); SSDiff surpasses DiffBIR, demonstrating generalization to pose variations. (DiffBIR has shown strong profile-view visual results in its paper.)

  Real profile-view	DifFace	DiffBIR	PGDiff	Ours
  FID($\downarrow$)	151.5	146.0	158.1	143.5
  NIQE($\downarrow$)	4.116	5.803	6.225	4.035
  MAN-IQA($\uparrow$)	0.3603	0.3912	0.3712	0.3955

  Hard cases: Some visualizations already come from hard-degradation subsets (e.g., first row in Fig. 1, the third row in Fig. 5, and the second row in Fig. 6 in our main paper). We understand that this question may refer to more extreme degradations. Therefore, we evaluate on the first 200 faces of the real benchmark Wider-Test (severely degraded but no breakages); SSDiff remains competitive.

  Severe degradation	DifFace	DiffBIR	PGDiff	Ours
  FID($\downarrow$)	82.60	82.75	91.80	81.02
  NIQE($\downarrow$)	4.403	5.891	4.930	4.092
  MAN-IQA($\uparrow$)	0.3386	0.3781	0.3415	0.3697

  Wearing glasses: Due to scarcity, we collect 10 pairs of old-photo faces (glasses vs. non-glasses) with similar degradation levels from the web; the performance gap is negligible, demonstrating that such cases do not affect the robustness.

  Glasses	Wear	Not wear
  FID($\downarrow$)	128.4	127.6
  MAN-IQA($\uparrow$)	0.3838	0.3814

  We further evaluate 100 faces with glasses from real benchmarks (LFW-Test, WebPhoto-Test, and Wider-Test), where SSDiff performs competitively with only a minor gap to DiffBIR, which is built upon a fine‑tuned SD 2.1 model with heavy training costs, while our method is training-free and inference on a frozen DDPM generator.

  Real data (wear glasses)	DifFace	DiffBIR	PGDiff	Ours
  FID($\downarrow$)	136.3	119.8	126.9	121.7
  NIQE($\downarrow$)	3.871	5.567	5.241	3.843
  MAN-IQA($\uparrow$)	0.3558	0.4020	0.3284	0.3962

  Finally, VintageFace mainly consists of front-facing photos because decades ago, portraits were usually taken in studios to clearly record one's appearance, so profile views are rare. Similarly, glasses were uncommon at that time, so faces with glasses are rare. Therefore, our results don't contain examples of this. We will add these results in the final version.

• Weak1 (c): Clarification on Fig. 8. All ablations, including those summarized in Fig. 8, are conducted on our Medium subset, which we consider representative because it spans a wide range of degradations from mild to near-severe. The single input image shown in the figure is only for illustration due to layout constraints and may have caused confusion. We will make this clear in the final version.

• Weak1 (d): Performance when eyes, nose, and mouth are with heavy scratches. As shown in Fig. 3 and in the second and third rows of Fig. 5 in the main paper, SSDiff is able to recover scratches around eyes, nose, and mouth with natural visual quality. To validate more heavy scratches, we select 50 moderate‑scratch faces and 50 severe‑scratch faces of these regions from VintageFace. We evaluate the mean smoothness (e.g., Edge strength) before and after restoration. The table shows that although severe scratches produce poorer smoothness in inputs of these regions, smoothness of restorations with no significantly degraded areas, demonstrating the robustness of our SSDiff for eyes, nose, and mouth. This robustness comes from generative priors of diffusion, which use contextual facial cues and pseudo-reference guidance to recover these regions. Visualization can also prove this, in Fig. 11 of the main paper, where even heavily degraded, eyes, nose, and mouth are recovered relatively well.

  Type	Medium scratches	Severe scratches
  Edge Strength($\downarrow$)	Input:114/Restored:46	Input:140/Restored:51

• Weak1 (e): Identity preservation. Identity consistency is preserved by combining the fidelity signal from a restorer with diffusion guidance. The restorer first produces a coarse but aligned structure and texture, and in reliable regions without breakages, this output is close to the original identity. These reliable regions are enforced through our L1 loss in all reverse diffusion processes, so that the guidance passed to diffusion models retains the original appearance, preventing identity drift in generation.

  we avoid explicit constraints like landmarks or 3D face shape because such constraints are highly sensitive to errors and even fail under moderate degradations. Instead, our design follows the same philosophy as generation-based restoration methods for controlling fidelity: using a reliable restoration module to preserve original facial textures as a constraint. However, the difference is that we then use diffusion guidance rather than training optimisation to refine images. This avoids the need for end-to-end training while effectively constraining the diffusion so that the results do not drift away from the true identity.

• Weak1 (f): Handling of background regions. The main reason for the less satisfactory background restoration is that the detected scratch fails to fully cover all breakage areas, and the diffusion itself has relatively weak generative ability for backgrounds compared to faces. Nevertheless, resulting backgrounds remain visually reasonable, showing the robustness of SSDiff for such detection error. We appreciate your observation, which indeed points out a potential direction for future work; we plan to explore region-adaptive guidance, where different areas (e.g., face vs. background) are guided with adaptive strengths to address this issue.

• Weak2 (a): Line 191-195. The current formulation may cause a misunderstanding. The intention of Eq. 9 was not to suggest that $s_w$ and $s_s$ are learned during network optimization. We wish to express that empirically determining these two coefficients by manual search can minimize the overall guidance loss, achieving optimal restoration. In the final version, we will revise Eq. 9 to make clear that these are fixed hyperparameters rather than parameters obtained through optimization.

• Weak2 (b): Robustness of Face Parsing. Our claim is based on extensive observations: under heavy degradations, facial landmarks and 3D face estimation often fail completely, whereas parsing networks still produce approximate facial contours. This is also supported by our analysis in "Weak2" of Reviewer pXcG, where the IoU value between severely degraded and GT facial contours is high (79%), and we further show that SSDiff remains robust even with imperfect parsing maps in that response. We will add more visual examples to better support this.

• Weak2 (c): Single-step use of style transfer. Style transfer is applied only once, at the moment when face structures become relatively stable ($T=T_1$). The color distribution obtained at this moment provides a coarse but reliable color reference, which is then propagated and refined by the diffusion during the remaining reverse steps. Because the Selective Restoration Guidance is applied throughout, this single-step color initialization doesn't harm structural fidelity. Repeating style transfer at every diffusion step offers little additional benefit while incurring prohibitive computational cost ($T_1$ times longer runtime), so a single application is enough.

• Weak2 (d) & Q1 & Q2: Unclear of $\hat{x}_{t}$. In Fig. 3, $\hat{x}_{t}$ denotes the intermediate restored estimate of clean images at timestep t in the reverse diffusion. Specifically, at each step, the model predicts a clean image from the current noisy state (as formulated in Line 7 of Algorithm 1 in our main paper), and $\hat{x}_{t}$ is this predicted clean sample. It indicates how images progressively become accurate as reverse diffusion proceeds. We will revise the figure caption in the final version.