acceptodds
Under review as a conference paper at ICLR 2027

InfinityFace: Unlocking Visual Autoregressive Text-to-Image Models for Blind Face Restoration

Abstract

Blind face restoration (BFR) is ill-posed, requiring a balance between perceptual realism and fidelity to the degraded input. We explore adapting Infinity, a bitwise visual autoregressive text-to-image model, to BFR. It is non-trivial as generation synthesizes plausible content from a generative prior, whereas restoration must retain the reliable input evidence. We propose InfinityFace, which reformulates open-ended next-scale generation as input-constrained next-scale restoration through two complementary forms of LQ conditioning. Specifically, a scale-dependent suffix controller combines the restored prefix with the corrected unresolved LQ features, preserving reliable LQ evidence while suppressing fine-scale feature mismatch. Meanwhile, scale-agnostic Text–LQ conditioning constructs an LQ visual memory and jointly injects it with text through multimodal cross-attention. Together, they anchor restoration to the observation while leveraging generative priors for realistic detail refinement. Experiments on synthetic and real-world BFR benchmarks show strong no-reference perceptual quality and competitive restoration performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
blind face restoration, visual autoregressive model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RegyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Well, InfinityFace adapt the Infinity bitwise VAR text-to-image model to blind face restoration, feeding a corrected LQ suffix at each scale and LQ visual features alongside the text in cross-attention. It leads MANIQA and TOPIQ on real-world sets but has the lowest CelebA-Test PSNR.

Critical strengths

Putting LQ information into the unresolved suffix fits bitwise next-scale prediction well and differs from VARSR's static LR prefix. Real-world result are good, best or second-best on every no-reference metric on LFW and Wider, and the user study gives it the highest Rank-1 rate on every split.

Critical weaknesses
  1. Identity is never measured, though CodeFormer and OSDFace report it. This matters because CelebA-Test PSNR is below the LQ input, and Fig. 9 shows text alone changes age texture.

  2. The ablation is thin. Table 3 is Wider-only with no-reference metrics, every variant keeps an LQ suffix, and suffix control alone lowers MUSIQ (74.02 to 73.54) though the text says all metrics improve.

  3. VARSR, the closest baseline, only has its VQ-VAE fine-tuned on FFHQ (line 318). If the rest stays a generic SR model, the gap mostly reflects face-specific training, not the proposed conditioning.

Questions for authors
  1. In Fig. 9 the glasses visible in the LQ input vanish once the caption drops them. How often do captions miss attributes, and can LQ evidence override the text?

  2. What are the latency and peak memory, including Qwen captioning, versus OSDFace and CodeFormer on the same GPU?

  3. Why re-synthesize CelebA-Test with your own degradation (line 297) rather than use a version from prior work?

8 comments
AuthorsOpenReview ↗

In response to Reviewer Regy’s questions (Q1)

We thank Reviewer Regy for the constructive feedback. We have made every effort to address the three questions (Q1–Q3) through additional experiments and analyses, manuscript revisions, and detailed point-by-point responses below. We welcome further questions and discussion.

Q1: (1) How often do captions miss attributes?

Caption audit. We selected 100 images at evenly spaced indices from the 3,000-image CelebA-Test set and conducted a careful manual audit of their Qwen-generated captions. For each sample, we inspected the caption alongside the LQ input and its corresponding GT image, reviewing each attribute to identify missing descriptions and assess the accuracy of explicit claims, using GT as the reference. The audit covered gender, age, skin tone, hair color, expression, and additional features such as earrings, mustaches, and eye makeup. The prompt permits descriptions of clearly visible attributes and excludes uncertain details.

Category definitions. We use five categories to evaluate captions generated from LQ images against the corresponding GT images. Correct denotes an explicit description supported by GT. Incorrect denotes a description that contradicts GT or makes an attribute claim without visible support in GT. Not mentioned denotes a predefined basic attribute absent from the caption; for example, person does not specify gender. Omitted denotes an additional feature visible in GT but absent from the caption. Not applicable denotes cases where GT contains no additional feature requiring description and the caption makes no additional feature claim.

AttributeCorrectIncorrectNot mentionedOmittedNot applicable
Gender961300
Age960400
Skin tone980200
Hair color917200
Expression8101900
Additional features57180421

Across the five basic attributes, captions left attributes unmentioned in 30 of 500 image–attribute pairs (6.0%): gender (3%), age (4%), skin tone (2%), hair color (2%), and expression (19%). Additional features visible in GT were omitted in 4 of 100 captions (4%). Missing descriptions were therefore most frequent for expression. The categories below distinguish caption completeness from description accuracy.

Correct descriptions of the five basic attributes range from 81 to 98 per attribute across the 100 captions. Incorrect descriptions occur only for gender (1) and hair color (7); no incorrect descriptions are recorded for age, skin tone, or expression. For additional features, 57 cases are labeled correct, 18 incorrect, and 21 not applicable, alongside the 4 omissions noted above.

(2) Can LQ evidence override the text?

Yes. Our results are consistent with LQ evidence taking precedence over text in some cases, although this is not guaranteed. A plausible explanation is that discernible visual cues constrain the restoration, whereas weak or ambiguous cues allow text to exert greater influence. As detailed below, the model can retain information absent from the caption and avoid some incorrect descriptions, but does not consistently prioritize LQ evidence over text. This comparison is descriptive and does not isolate the causal contribution of LQ conditioning.

We have added face examples and accompanying discussion in Appendix Figure 12 of the revised manuscript:

  • Column 1: retaining information omitted from the caption. An earring omitted from the caption is retained with a location and appearance consistent with GT, demonstrating recovery of information beyond the caption.
  • Column 2: omission combined with a conflicting textual description. The caption omits a cap and instead describes short dark hair; the output replaces the cap with hair, consistent with the conflicting textual cue.
  • Column 3: avoiding an incorrect textual description. The incorrectly described chin-on-hand pose is not reproduced, demonstrating resistance to erroneous text.
  • Column 4: reproducing an incorrect textual detail. The restoration reproduces the incorrect earring description, showing that an erroneous textual detail can also appear in the output.

In the inspected outputs, most failures of LQ evidence to override text occur for incorrect or omitted additional features, whereas incorrect descriptions of the first 5, coarser-grained attributes are usually overridden. Across the 4 omitted-feature images, 2 retain the omitted feature, 1 partially retains it, and 1 loses it. None of the five corresponding restorations shows the incorrectly described chin-on-hand pose. Of the 9 cases involving incorrect earring descriptions, 4 visibly generate earrings, 4 show no visible added earrings with the ears obscured by hair, and 1 remains ambiguous. The 1 annotated gender-description error is also not corrected, although the LQ image provides ambiguous visual cues.

AuthorsOpenReview ↗

In response to Reviewer Regy’s questions (Q1-Q3)

Role of degradation. Our fixed-caption degradation analysis suggests that discernible LQ cues can support recovery of attributes absent from or in conflict with the caption, while weaker cues leave more room for text and the generative prior to influence the output.

In Section 5.4 and Figure 8 of the revised manuscript, we examine two examples across four degradation levels from mild to severe, keeping one caption fixed for each example. In the upper example, the caption omits a visible cap and instead describes short dark hair. The first two restorations retain a recognizable cap despite this description; at the heaviest degradation, short dark hair replaces the cap, consistent with the caption. In the lower example, the caption omits facial expression and mouth state. The first three restorations retain a tooth-bearing smile, while the heaviest degradation produces a closed-mouth smile: the broad expression persists, but the visible teeth are lost. This behavior depends on the visibility of the particular attribute in the LQ image; these examples do not establish a universal degradation threshold at which text takes precedence.

Interpretation of Figure 9 in the original paper (now Figure 13 in the revised manuscript). The changes in glasses and age-related texture across captions illustrate how text can influence the restored appearance for the same LQ input. The purpose of this figure is to demonstrate a possible use of this capability: when severe degradation makes an attribute difficult to discern, reliable information about the subject from another source could inform the caption and guide restoration toward the subject's known characteristics.

Q2: Inference time and peak GPU memory.

We report per-image inference latency and peak allocated GPU memory for the compared methods under fixed single-GPU evaluation protocols.

MethodLatency (s)Peak memory (GiB)
CodeFormer0.0370.82
DAEFR0.0511.43
DifFace2.5991.03
OSDFace0.1424.88
VARSR0.2998.13
InfinityFace0.69818.08
Qwen captioning + InfinityFace1.73634.42

With an existing caption, InfinityFace includes T5 encoding and restoration, taking 0.698 s/image with 18.08 GiB peak memory. Including online Qwen captioning increases latency to 1.736 s/image and peak memory to 34.42 GiB with all models co-resident.

Q3: Why synthesize CelebA-Test inputs rather than use a released version?

Like CodeFormer and DifFace, we synthesize our own test inputs. For CelebA-Test, we use the Real-ESRGAN pipeline. All methods are evaluated on the same synthesized inputs and the same real-world benchmarks.

Thank you for the constructive questions and we will address the concerns raised in the Weaknesses section in a separate follow-up response.

Okay, I'll wait for your weakness.

Reviewer_3vvFOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on the problem of blind face restoration. Specifically, the authors propose a method called InfinityFace that is built on the pretrained Infinity text-to-image model. The core insight of this method is to reformulate next-scale generation as restoration by combining previously restored scales with unresolved low-quality (LQ) features. In the proposed InfinityFace, the two main components are a scale-dependent LQ suffix controller that adjusts these features and a scale-agnostic conditioning mechanism that jointly incorporates LQ visual memory and text through cross-attention.

Critical strengths
  1. This work adopts the AR model into the blind face restoration task. It sounds interesting.

  2. The HQ-prefix oracle analysis shows that unresolved LQ features provide useful coarse-scale guidance but can introduce mismatch at finer scales. This analysis provides a concrete motivation for scale-dependent suffix control and strengthens the rationale for the proposed design.

  3. InfinityFace achieves strong no-reference quality scores on synthetic and real-world benchmarks. The ablation studies show that the two conditioning mechanisms complement each other. The user study also shows that participants prefer its results, especially on real-world inputs. These findings support its value for improving the perceptual quality of restored faces.

Critical weaknesses
  1. The work presents the method as a reformulation from generation to restoration. However, the main changes introduce LQ conditioning into an existing pretrained model. Combining a restored prefix with an unresolved LQ suffix is a reasonable design, but comparisons with simpler conditioning methods using the same backbone are needed to establish its value. In addition, the paper lacks experiments on other pretrained models to assess the generalizability of the proposed approach.

  2. The suffix controller uses an MLP to adjust features, while the conditioning mechanism uses cross-attention to incorporate image and text features. The novelty mainly lies in their combination and placement, which limits the methodological contribution.

  3. Figure 3 shows that raw LQ suffixes can introduce feature mismatch at finer scales. However, the paper does not directly show whether the controller reduces this mismatch or how its behavior changes across scales. This leaves a gap between the motivation and the demonstrated effect of the controller.

  4. The main advantages are in no-reference quality metrics. Reference-based results on the synthetic dataset are less competitive, and no face identity similarity metric is provided. These results do not sufficiently establish whether the method preserves identity while improving visual quality.

Questions for authors
  1. Have you compared the proposed method with simpler LQ conditioning strategies using the same Infinity backbone, such as cross-attention alone without LQ suffix input?

  2. Does the suffix controller reduce the fine-grained feature mismatch identified in Figure 3?

  3. What are the inference time and peak GPU memory usage compared to the compared methods?

  4. Have you tested the proposed design with another pretrained backbone or Infinity model size?

6 comments
AuthorsOpenReview ↗

In response to Reviewer 3vvF’s questions (Q1–Q4)

We thank Reviewer 3vvF for finding our work interesting and for the constructive feedback. We have made every effort to address the four questions (Q1–Q4) through additional experiments and analyses, manuscript revisions, and detailed point-by-point responses below. We welcome further questions and discussion.

Q1: Comparison with simpler LQ conditioning.

Comparison setting. We evaluated a simpler conditioning strategy using the same Infinity backbone, retaining text conditioning and introducing LQ information only through cross-attention, without an LQ suffix. This was also the approach we explored at the start of our research. We have added its visual results as the "No Suffix" setting in Figure 6.

Results and motivation. Without the LQ suffix, the model reproduces only some input characteristics, such as the overall color palette and certain facial attributes, while showing poor fidelity to the input face. This observation motivated our reformulation of next-scale prediction as next-scale restoration: combining the restored prefix with the unresolved LQ suffix keeps spatially aligned input evidence available at each scale. The comparison supports the role of the suffix in anchoring restoration to the input face. We discuss the scope of backbone validation in Q4 below.

Q2: Whether the controller reduces the feature mismatch in Figure 3.

We have added a comparison of the visual states formed by the restored prefix combined with either the raw or the controlled LQ suffix. Newly added Figure 7 shows lower RMS distance and higher cosine similarity with control at scales 7--10, with small differences at scales 2--6, supporting improved fine-scale alignment during restoration.

Q3: Inference time and peak GPU memory.

We report per-image inference latency and peak allocated GPU memory for the compared methods under fixed single-GPU evaluation protocols.

MethodLatency (s)Peak memory (GiB)
CodeFormer0.0370.82
DAEFR0.0511.43
DifFace2.5991.03
OSDFace0.1424.88
VARSR0.2998.13
InfinityFace0.69818.08

With a provided caption, InfinityFace takes 0.698s per image (including T5 encoding and restoration, excluding Qwen captioning) with 18.08 GiB peak memory.

Q4: Evaluation with other backbones and model sizes.

Exploration scope. We thank the reviewer for suggesting this direction. Our current study focuses on Infinity-2B to investigate how its native text-to-image prior can support blind face restoration. We have not evaluated the restoration design with another pretrained backbone or completed restoration experiments with another Infinity model size, so its generalizability across backbones and model sizes remains unverified.

Other model sizes. Among the official checkpoints we considered, the only smaller option was a 125M model designed for 256 × 256 image generation. We conducted a preliminary face-domain adaptation trial with this checkpoint. Its generation quality after adaptation was not satisfactory, so we did not proceed to restoration experiments. The larger 8B and 20B models exceed our available computational resources. We consider evaluation across additional backbones and model scales a useful extension of this study.

Thank you for the constructive questions and we will address the concerns raised in the Weaknesses section in a separate follow-up response.

AuthorsOpenReview ↗

In Response to Reviewer 3vvF's Weaknesses(W1-W4)

Following our submitted responses to your questions, we address the concerns in the Weaknesses section below.

W1: Comparison with simpler LQ conditioning and generalizability.

We have addressed these concerns in our submitted responses to Q1 and Q4. Please refer to Q1 for the simpler LQ-conditioning comparison using the same backbone, and Q4 for the scope of evaluation across pretrained backbones and model sizes.

W2: Methodological contribution and design rationale.

We agree that MLPs and cross-attention are established components. Our contribution lies in the restoration formulation and the complementary roles assigned to LQ information within it.

Complementary roles. The LQ suffix forms part of the autoregressive visual state, supplying dense, spatially aligned evidence alongside the restored prefix. LQ visual memory supplies information through cross-attention that the transformer can read when refining its hidden representations. Please refer to our submitted response to Q1 for the no-suffix comparison that examines the role of the suffix.

Design rationale and evidence. We use a simple MLP to correct the unresolved suffix in VAE feature space, and cross-attention to make LQ visual memory accessible through Infinity's native conditioning interface. These choices follow the two roles above and remain compatible with the pretrained backbone. Table 3 reports the component ablations and their effects on perceptual quality, while Figure 6 provides visual comparisons. The controller's feature-space behavior is addressed in our submitted response to Q2.

W3: Whether the controller reduces the feature mismatch in Figure 3.

We have addressed this concern in our submitted response to Q2. Please refer to that response and Figure 7 of the revised manuscript for the comparison across scales.

W4: the relation between identity fidelity and visual quality.

Our quantitative results, including the added ArcFace IDS against GT, reflect a clear trade-off between fidelity and visual quality. Among the restoration methods, CodeFormer achieves the strongest pixel-level fidelity (PSNR/SSIM) and the highest identity fidelity (IDS) apart from OSDFace, which uses explicit identity supervision. However, CodeFormer is less competitive on no-reference perceptual metrics. This trade-off is also evident in DAEFR, which achieves higher MUSIQ/MANIQA scores but lower PSNR/SSIM/IDS than CodeFormer.

Following this principle, our InfinityFace outperforms CodeFormer on CLIPIQA, MUSIQ, MANIQA and TOPIQ, achieving the best MUSIQ and MANIQA scores among the compared methods. More importantly, in the user study, which participants had access to GT references and were instructed to jointly consider fidelity and perceptual quality on CelebA-Test, it also received the highest user preference rates. Full user study across real-world datasets further supports its ability to produce realistic, plausible facial textures and a coherent overall appearance.

MethodPSNR↑SSIM↑LPIPS↓IDS↑MUSIQ↑MANIQA↑TOPIQ↑FID↓CLIPIQA↑
LQ input25.330.7060.5740.81717.000.1600.198145.230.296
CodeFormer25.150.6850.2270.84775.550.5380.86453.100.671
DAEFR22.590.6270.2430.76575.280.5440.85440.640.668
DifFace24.800.6850.2720.74168.960.4680.77838.320.571
OSDFace23.380.6580.2400.87475.460.5500.85344.180.695
VARSR23.770.6810.2610.83275.040.5120.86847.870.604
InfinityFace22.590.6440.2660.74175.570.5720.89148.560.681

Thank you again for the constructive feedback.

Dear Reviewer 3vvF,

As the discussion period is nearing its end, we would like to follow up on our responses to your questions and weaknesses. We hope these responses and the additional results have addressed your concerns. If you have any remaining questions or concerns, we would appreciate the opportunity to discuss them and provide further clarification.

Thank you again for your time and thoughtful feedback. If our responses have addressed your concerns, we would be grateful if you could reconsider your overall assessment and score.

Best regards,
The Authors

Reviewer_hVM5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adapts Infinity (a bitwise VAR text-to-image model, BSQ + coarse-to-fine next-scale prediction) to blind face restoration (BFR). The key reframe: turn open-ended "next-scale generation" into input-anchored "next-scale restoration". An oracle analysis shows raw LQ suffixes help coarse structure but mismatch fine scales. So it adds two things. (1) A scale-dependent LQ suffix controller that corrects the LQ suffix in VAE space, conditioned on the restored prefix. (2) Scale-agnostic Text–LQ conditioning: build an "LQ visual memory" and inject it with text via shared-query cross-attention.

Critical strengths

Likely the first to adapt a bitwise VAR T2I prior (Infinity) to BFR, and the scale-dependent conditioning is grounded in a clean oracle analysis, not just asserted.

Critical weaknesses
  • Modest gains, heavy method. Not best on FID/CLIPIQA and among the worst on PSNR/SSIM (CelebA); best only on MANIQA/TOPIQ on real sets. For a 2B, 10-scale AR pipeline, the gain over a one-step baseline like OSDFace is incremental, with unreported cost.
  • Text branch may be unnecessary. Caption-source ablation (Table 4) shows LQ / HQ-oracle / fixed prompt differ only slightly (FID 48.56 / 47.05 / 49.08). The paper never isolates text vs LQ-memory to show text is actually needed.
Questions for authors

Report inference latency / FLOPs / params vs baselines (CodeFormer, OSDFace, VARSR). What is the quality–latency trade-off?

4 comments

Response to Reviewer hVM5

We thank Reviewer hVM5 for the insightful feedback. We have added experiments and analyses, revised the manuscript, and provided point-by-point responses below. We welcome further questions and discussion.

W1 and Q1: Performance gains, computational cost and the quality–latency trade-off.

We thank the reviewer for requesting a clearer quantitative comparison of the performance gains and computational requirements. We address these aspects below.

Performance gains. InfinityFace ranks first or second across all five evaluated metrics on both LFW-Test and Wider-Test, achieving the highest MANIQA and TOPIQ scores on both datasets. Compared with OSDFace, TOPIQ improves by an absolute 0.041 on each dataset. These gains are further supported by the highest user-study first-choice rates of 38.25% and 40.35%, respectively. On CelebA-Test, InfinityFace achieves lower PSNR/SSIM scores while delivering strong perceptual restoration quality. This emphasis on perceptual realism aligns with recent methods such as OSDFace, VARSR, and GeoMAR, which also exhibit a trade-off between pixel-level fidelity and perceptual quality. CodeFormer achieves the best pixel-level fidelity in our evaluation while remaining less competitive in perceptual quality, further illustrating this trade-off.

Computational cost. As requested, we report inference latency, parameter count, peak allocated GPU memory, and covered FLOPs under fixed single-GPU protocols, using compared methods as baselines. The parameter count for InfinityFace includes a 2500.57M backbone, a 110.12M VAE, and a 1223.53M T5 text encoder. The final row additionally includes online Qwen captioning.

MethodParameters (M)Latency (s)Peak memory (GiB)FLOPs (G)
CodeFormer94.110.0370.82809.77
DAEFR152.150.0511.43906.77
DifFace175.582.5991.0337372.19
OSDFace979.130.1424.884635.19
VARSR1210.860.2998.1314142.82
InfinityFace3834.220.69818.0821661.82
Qwen captioning + InfinityFace12601.341.73634.42—

Quality–latency trade-off. InfinityFace restores an image in 0.698s, increasing to 1.736s with online Qwen captioning. Compared with OSDFace’s 0.142s/image, this higher inference cost comes with measurable perceptual gains on both real-world benchmarks, supported by the user study. These results position InfinityFace as an approach that prioritizes perceptual quality for real-world face restoration and support the effectiveness of multi-scale autoregressive restoration when perceptual quality is the primary objective.

W2: Whether text is needed beyond LQ visual memory and suffix control.

Isolating text conditioning. The ablation in the original Table 3 retained text conditioning in all settings. To directly assess this contribution, we added a no-text variant (e) that preserves both LQ visual memory and suffix control. Comparing the full model (d) with variant (e) isolates the effect of text conditioning. We have updated Figure 6 and Table 3 to explicitly identify (e) as "w/o text."

Results. Compared with no-text variant, the results drop in all metrics on LFW and Wider. Qualitatively, on the same examples used in Figure 6, the model trained without text and identity conditioning exhibits unnatural periocular and dental structures. In row 1(e), both the eyeballs and the mouth exhibit pronounced structural distortions and unnatural artifacts. In row 2(e), the sunglasses lose their coherent frame and lens structure, leaving dark regions intermingled with the generated eyes. These examples show that sharp local textures alone do not ensure coherent facial structures. Together, the quantitative and qualitative comparisons support the complementary contribution of text conditioning to perceptual quality and local structural coherence.

LFW-Test

MethodFID↓CLIPIQA↑MUSIQ↑MANIQA↑TOPIQ↑
(d) Full model46.650.70975.830.5540.893
(e) w/o text47.760.68674.520.5230.850

Wider-Test

MethodFID↓CLIPIQA↑MUSIQ↑MANIQA↑TOPIQ↑
(d) Full model36.790.72475.130.5520.880
(e) w/o text42.730.70073.190.5150.846

Thank you again for the constructive feedback on computational cost and the role of text conditioning.


Dear Reviewer hVM5,

As the discussion period is nearing its end, we would like to kindly follow up on our responses, which we submitted a week ago. We would greatly appreciate it if you could review them and let us know whether they adequately address your concerns. We would be happy to provide any further clarification.

Thank you again for your time and thoughtful feedback. If our responses have addressed your concerns, we would be grateful if you could reconsider your overall assessment and score.

Best regards,
The Authors

AuthorsOpenReview ↗

In response to Reviewer hvM5's weakness 2

Our earlier response reported the no-text comparison on LFW-Test and Wider-Test, where the full model performs better on all five evaluated metrics. To supplement that response, whose length was constrained, we provide the CelebA-Test results here. The full model performs better on five of eight metrics.

MethodPSNR↑SSIM↑LPIPS↓FID↓CLIPIQA↑MUSIQ↑MANIQA↑TOPIQ↑
Full model22.590.6440.26648.560.68175.570.5720.891
w/o text22.960.6470.28247.450.66074.080.5380.854

Dear Reviewer hVM5,

As the discussion period is nearing its end, we would like to follow up on our previous responses and summarize how your feedback has helped deepen the analysis in our paper.

  • For W1/Q1, we supplemented the performance comparisons and computational cost measurements, clarifying the quality–latency trade-off.

  • We particularly appreciate your suggestion to examine the contribution of text conditioning separately. In response, we added the section “Effect of Removing Text Conditioning”, with quantitative and qualitative analyses of setting (e), w/o text, which retains both LQ visual memory and suffix control. The full model performs better on all five evaluated metrics on LFW-Test and Wider-Test, and on most evaluated metrics on CelebA-Test. More notably, despite producing sharp local textures, the no-text variant exhibits distorted eyes and mouth structures and disrupted sunglasses in Figure 6. These observations suggest that text helps the model exploit its pretrained generative prior to produce coherent facial structures when LQ evidence is ambiguous.

  • Your suggestion also motivated a deeper investigation of the relationship between text and LQ evidence. In Section 5.4 and Appendix A.1, we examine restoration across degradation levels with fixed captions, and vary text attributes while keeping the LQ input fixed. These experiments suggest that discernible LQ cues can support recovery of attributes absent from or conflicting with the caption, while weaker cues leave more room for text and the generative prior to influence the output. Together, these additions deepen the paper’s exploration of when and how text contributes to restoration.

We sincerely appreciate your constructive feedback and hope you will consider these clarifications and additional analyses when reassessing our work. We would be happy to clarify any remaining questions.

Thank you again for your time and thoughtful feedback.

Best regards,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Contributions and Rebuttal Updates

Dear Area Chair,

We sincerely thank you and Reviewers hVM5, 3vvF, and Regy for the careful evaluation of our work. We summarize the main additions and clarifications below for your consideration. To make these changes easy to locate, first-round revisions are marked in red and second-round revisions in blue.

In addition to clarifying the method, evaluation, and scope of our claims, we strengthened the experimental evidence and extended our analysis in three main directions:

  1. Strengthened ablations and an additional adapted baseline.
  • We added a no-suffix comparison and setting (e), w/o text. The revised manuscript presents the Wider-Test quantitative ablation in Table 3 and the visual comparisons in Figure 6, with a dedicated “Effect of Removing Text Conditioning” analysis. The full model performs better on all five evaluated metrics on LFW-Test and Wider-Test and on most evaluated metrics on CelebA-Test; please refer to our response to Reviewer Regy's W2 for the expanded ablation tables across all three datasets. Visually, despite producing sharp local textures, the no-text variant exhibits distorted eyes and mouth structures and disrupted sunglasses, providing evidence that text contributes to coherent facial structures beyond texture sharpness.
  • We also strengthened the baseline comparisons by further fine-tuning VARSR on FFHQ using the same degradation protocol as InfinityFace. We incorporated this adapted baseline into the main quantitative and qualitative comparisons and reported both the original and updated results across all three datasets.
  1. Deeper understanding of suffix control and text–LQ interaction.
  • We added a direct feature-space analysis of suffix control, showing lower RMS distance and higher cosine similarity to the full-HQ reference at finer scales. These results provide evidence for the controller’s role in reducing input-state mismatch during restoration.
  • We also expanded Section 5.4 and Appendix A.1 to examine fixed-caption restoration across degradation levels, text interventions on fixed LQ inputs, and a manual caption audit on 100 images with analysis of the corresponding restorations. These experiments suggest that discernible LQ cues can support the recovery of attributes absent from or conflicting with the caption, whereas weaker cues leave more room for text and the generative prior to influence the output. Together, these analyses deepen our understanding of how the conditioning mechanisms operate and when text contributes to restoration, while helping address Reviewer hVM5’s concerns about text conditioning.
  1. Initial exploration of applicability across model sizes during the final hours of discussion.
  • In the final hours of the discussion period, we continued responding to the reviewers’ suggestions by investigating the applicability of our method to a smaller model. Specifically, we completed a preliminary training run with 24 transformer blocks instead of 32, using five RTX 4090 GPUs, a total batch size of 25, and 500 iterations, corresponding to 12,500 training image instances. This experiment represents an initial step toward assessing applicability across model sizes under a limited training budget; further training and evaluation will follow.

Positive reviewer reassessment. We are encouraged that Reviewers 3vvF and Regy, the two reviewers who participated in the follow-up discussion, acknowledged that most of their concerns had been addressed and raised their scores to 3. We appreciate their recognition of these improvements and their suggestions for further refinement. Reviewer hVM5’s comments focused on requests for clarification rather than concerns about our method or its novelty. We provided detailed responses and follow-up summaries, including computational comparisons and expanded analysis of text conditioning. Although we have not yet received a follow-up response from Reviewer hVM5, we are confident that our clarifications and additional evidence address the concerns they raised.

We believe that the methodological contribution, strengthened experimental evidence, and deeper analysis demonstrate the value of our work and merit consideration for acceptance. We remain committed to incorporating the reviewers' suggestions into the next revision, completing the reduced-depth experiments, exploring additional frameworks such as VAR and Switti, and addressing the remaining limitations and any issues identified during further validation. We sincerely thank the reviewers for helping improve the paper and appreciate your consideration.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.