acceptodds
Under review as a conference paper at ICLR 2027

EMBER: Rekindling Old Photo Restoration from Early Trajectories and No-Reference Feedback

Abstract

A faded photograph can be the last visible trace of a person or moment that no longer exists, but restoring it is difficult without clean references or reliable damage masks. Our approach begins with two observations about a pretrained image editor. Its first two denoising steps produce a spatial edit-response signal associated with the regions where the final output later differs from the input. No-reference policy optimization also amplifies restoration and colorization responses that supervised fine-tuning does not consistently express. Based on these observations, we introduce EMBER, which combines a Gated 1–2 Step Edit-Response Field (G12) with no-reference group-relative policy optimization (GRPO) on real old photographs. G12 freezes the early response into mask-free rotary-phase modulation, while GRPO adapts the editor using a reward with no explicit color statistic. Across GGT-100K Old Photo and RealOld-74, EMBER performs best or remains competitive on most no-reference metrics not used as training rewards. In a blinded study, 31 participants completed 1,018 valid image-level evaluations across 48 photographs, selecting their two preferred results in each evaluation. Across all seven methods, EMBER achieved the highest Top-2 inclusion rate at 58.3%. EMBER restores these photographs in the form in which they often survive: as a single damaged image, without clean targets during policy optimization or external masks at deployment. Code is available at https://anonymous.4open.science/r/artifact-af5213a/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Old photo restoration, Image Restoration, Diffusion Models, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_H8s8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets real old-photo restoration from a single damaged image, without clean references, degradation masks, or paired real supervision. It builds on a frozen Qwen-Image-Edit-2511 backbone with LoRA, and proposes two components: G12, which extracts an edit-response field from the first two denoising steps and freezes it as a target-token RoPE phase modulation for later steps; and no-reference GRPO, which adapts the editor on unpaired real photos using MUSIQ + CLIP-IQA group-relative rewards plus an SFT-anchor penalty. Claims are supported by no-reference metrics on GGT-100K OldPhoto and RealOld-74, plus a blinded 7-method human study.

Critical strengths
  1. Practical and well-scoped problem setting.​ Single-image real old-photo restoration without masks or paired targets is the right deployment assumption; the paper avoids over-relying on synthetic paired evaluation.
  2. G12 is conceptually interesting.​ Using early-trajectory response as an internal, mask-free spatial signal is a reasonable alternative to DiffEdit-style prediction difference or explicit mask conditioning.
  3. Honest about limitations.​ The authors explicitly state the field is not a damage map, that GRPO color response is not historical-color recovery, and that no-reference metrics are not independent confirmations. This is better self-calibration than many generative-restoration papers.
  4. Human preference study is a real asset.​ 31 participants, 1,018 ballots, Top-2 protocol, clustered bootstrap and conditional logit are unusually careful for this venue. EMBER’s 58.3% Top-2 rate is a meaningful signal.
  5. Engineering transparency.​ Pseudocode, field normalization, κ schedule, GRPO ratio/clipping, hardware, and bootstrap seeds are documented in unusual detail.
Critical weaknesses
  1. Incrementality of the core method.​ G12 combines early-step residual extraction [DiffEdit-like], RoPE phase offset [LooseRoPE-like], and frozen-field modulation; GRPO-on-LoRA reuses standard group-relative policy optimization . The paper reads as a competent adaptation of existing pieces rather than a new mechanism. The novelty claim should be narrowed to “early-trajectory modulation for mask-free restoration” and not overstated as a representation-level contribution.
  2. Weak causal evidence that G12 placement helps.​ The permutation intervention (Table 11) shows no significant damage-LPIPS gain from correct vs permuted field, and even a negative damage-L1 effect. The authors explain this away as “mask-defined endpoints don’t capture fading/color,” but that weakens the claim that G12 is doing useful spatial localization. The main G12-vs-non-G12 gain is also not a controlled ablation: the two rows come from different training trajectories/checkpoints, so the delta is end-to-end, not field-causal.
  3. Reward metrics are biased toward the method’s failure mode.​ MUSIQ/CLIP-IQA are known to reward high-frequency content, over-sharpening, and generative “pop” without fidelity . EMBER explicitly increases saturation/colorfulness and drops PSNR/SSIM/LPIPS after GRPO (Table 9), which is exactly the regime where these NR metrics inflate. Claiming “best or competitive on most no-reference metrics” is therefore circular unless balanced by human study or fidelity constraints — which the paper partially does, but the metrics are still over-used in the abstract.
  4. No-reference metrics are not “held-out” in practice.​ Four of five are learned IQA models that may share training-data/feature bias with CLIP-IQA/MUSIQ. The paper acknowledges this but still titles them “held-out NR metrics.” This wording overstates independence.
  5. Color recovery claim is vague by design.​ The instruction permits “natural-looking colors,” GRPO amplifies color response, but no historical-color check exists and arguably cannot exist. That is acceptable, but the paper should not imply color restoration quality beyond plausibility. The ablation also shows CLIP-IQA-only gives more color response, so the dual-IQA choice is not clearly justified except by monitor peaks.
  6. Full-reference degradation is large and under-discussed.​ On GGT-100K, EMBER(G12) gets PSNR 17.94 vs HYPIR 24.39, SSIM 0.69 vs 0.76. The paper frames this as perception-distortion tradeoff, but for old photos with identifiable people/places, identity and content fidelity matter. No face-ID retrieval, no structure-preservation metric, no “intact region” human rating.
  7. Human study stratification is thin.​ 48 photos, 4 per stratum, person/color/severity by visual inspection. Strata-level rates (Fig 12) are descriptive only; RealRestorer beats EMBER in color-present strata. The overall Top-2 win is convincing, but “EMBER restores old photos better” is stronger than the data supports for color photos.
  8. G12 field may encode “where model edits,” not “where damage is.”​ Correlation ρ≈0.31 partial is modest. Applying modulation on permuted fields barely changes LPIPS suggests the field acts more like a soft attention prior than a damage prior. The paper should clarify whether G12 is claimed to help localization or only trajectory steering.
Questions for authors
  1. Causal G12 ablation.​ Since EMBER(non-G12) and EMBER(G12) are different checkpoints/trajectories, can you report a same-anchor ablation where only the frozen field is enabled/disabled? Without it, Table 1 cannot attribute gains to G12.
  2. Field permutation result.​ Table 11 shows permuted field ≈ correct field on LPIPS and better on damage-L1. Does this mean G12’s main benefit is global phase offset rather than spatial placement? If so, is the 12×12 field unnecessarily complex?
  3. Reward bias.​ Given MUSIQ/CLIP-IQA reward sharpness/color , have you checked whether disabling the color instruction or using a fidelity-penalized reward (e.g. LPIPS-to-input or ID similarity) changes rankings?
  4. Leakage during SFT.​ The detached two-step preview avoids using x_s, but the condition latent z_y comes from the LQ image. Could the field still exploit dataset-specific LQ statistics that don’t transfer to real photos? Any real-photo field visualization?
  5. Identity/fidelity evaluation.​ Old photos often contain faces. Can you add ArcFace ID retention or a human “content preserved” rating, especially since GRPO lowers SSIM?
  6. Why κ=8 then 4?​ The taper is intuitive but unmotivated. Is there a sensitivity curve? Table 14 shows 1–4 observations are nearly identical — could G12 be simplified to a single-step global gate?
  7. Human study baselines.​ SeedEdit / Qwen-Edit are weak restoration baselines; the real competitor is RealRestorer and HYPIR. In color-present strata RealRestorer wins — do you see EMBER over-colorizing?
  8. Deployment cost.​ 8-step rollout + field estimation + LoRA is not cheap. Latency table uses synthetic 512×512; what is real-photo 1024px+ latency and memory? Is G12 actually “no extra denoiser pass” inclusive of the step-1–2 preview?
  9. Color correctness scope.​ Could you add a synthetic old-photo test with known ground-truth color to show whether GRPO color response correlates with correct color at all, or only “saturated natural-looking” output?
  10. Novelty statement.​ If DiffEdit, LooseRoPE, and Flow-GRPO already exist, what is the one claim that survives a “prior-work minus EMBER” test? I’d suggest reframing contributions around trajectory-derived mask-free modulation for restoration editors rather than generic “diffusion + GRPO.”
7 comments

We thank the reviewer for the detailed assessment across methodology, causal evidence, reward bias, and metric independence.

[W1] Incrementality of the core method

G12 combines DiffEdit-like residual extraction, LooseRoPE-like phase offset, and standard GRPO. The novelty claim should be narrowed.

Response. We have narrowed the claim to restoration-specific early-trajectory modulation and its field-replayed optimization protocol. G12 constructs an online edit-response field from early clean-latent estimates under one restoration instruction, gates temporal agreement, freezes the field, and applies it to later target-token frame-axis phases. We do not claim a new residual operator, rotary-embedding family, or policy optimizer.

Revised locations. Section 1 (p. 2); Appendix A (p. 16); Appendix F (p. 21).

[W2] Weak causal evidence; negative intervention result

Table 11 shows no significant LPIPS benefit from correct vs permuted field, and a negative damage-L1 result. The G12-vs-non-G12 comparison is not a controlled ablation.

Response. We retain the negative result. The original Table 11 is unchanged: neither primary LPIPS endpoint is significant; damage-L1 favors permutation (effect −3.128, 95% CI [−5.233, −1.229], p = 0.0021). New matched controls complement without reversing this.

Matched training (Table K.1, p. 29) — four arms, identical init/data/noise/budget:

Field (GRPO)Damage-LPIPS ↑Intact LPIPS ↓TOPIQ-NR ↑
Correct0.04680.08890.579
Permuted0.04410.09070.568
Constant0.04350.09140.563
Disabled0.04270.09220.556

Seed differences (correct − permuted): 0.0024 / 0.0028 / 0.0029 (mean 0.0027). Constant field cannot exclude all global-phase explanations (phase mapping is quantized).

Fixed-weight causal-path probe (Figure 5, p. 8; Appendix K.3, p. 30). Holding M-20 weights, noise, and all other cells fixed, changing only single cell phases: attention density change 0.04779 [0.04084, 0.05473], positive in all 20 photographs. All 16 phase-boundary crossings change reconstruction; all 4 same-bin controls produce zero change. This establishes a causal path through deployed RoPE without attributing quality improvement to it.

Revised locations. Appendix K.1 (p. 27); Table K.1 (p. 29); Figure 5 (p. 8); Appendix K.3 (p. 30); Table K.3 (p. 31).

[W3] Reward metrics biased toward the method's failure mode

MUSIQ/CLIP-IQA reward sharpness and color pop. Claiming "best on most NR metrics" is circular without fidelity constraints.

Response. The revised abstract pairs NR metric results with independent-set human preference and adverse fidelity findings. Controlled 0.75×→1.25× chroma contrast: CLIP-IQA +0.0061 [+0.0037, +0.0084], MUSIQ +0.31 [−0.06, +0.68]. Full-reference degradation remains explicit. Added identity/structure reward terms improve ArcFace but reduce TOPIQ-NR from 0.579 to 0.572. Human fidelity ratings shift adversely. We do not present these as eliminating reward exploitation.

Revised locations. Appendix H.1 (p. 24); Table H.1 (p. 25); Table N.3 (p. 42); Section 4 (p. 10).

[W4] "Held-out NR metrics" overstates independence

Four of five non-reward NR metrics are learned and may share bias with the reward models.

Response. We replaced "held-out NR metrics" with "non-reward NR metrics" throughout. The Table 1 caption now names the excluded rewards, states that four of five non-reward metrics are learned and may share feature or dataset biases with MUSIQ and CLIP-IQA, and identifies NIQE as handcrafted.

Revised locations. Table 1 (p. 5); Section 4 (p. 10).

[W5] Color recovery claim is vague; dual-IQA choice unclear

No historical-color check. CLIP-IQA-only gives more color response than dual-IQA, so the dual choice is not clearly justified.

Response. Real-photo claims are limited to plausible color response. Dual-IQA is treated as an evaluated operating choice, not a proven optimum. Three-seed evaluation: TOPIQ-NR 0.579 / 0.561 / 0.570 for dual-IQA / MUSIQ-only / CLIP-IQA-only. On 200 synthetic known-color sources, mean CIEDE2000 decreases from 8.9 (M-SFT) to 7.6 (M-20), supporting synthetic color improvement, not historical correctness on real photographs.

Revised locations. Table 5 (p. 23); Table H.1 (p. 25); Table L.1 (p. 36).

[W6] Full-reference degradation is large and under-discussed

GGT-100K: PSNR 17.94 vs HYPIR 24.39, SSIM 0.69 vs 0.76. No face-ID or structure-preservation metric.

Response. The gap is retained and now accompanied by direct fidelity measurements. M-20 vs M-SFT: −2.04 dB PSNR, −0.0725 SSIM, +0.0493 LPIPS.

MeasureM-SFTM-20Δ
Intact-region LPIPS (200 D-Syn) ↓0.08810.0906+0.0025
FFHQ ArcFace (100 sources) ↑0.8240.816−0.008
Real-photo ArcFace (42 valid) ↑0.7420.733−0.009 [90% CI: −0.0161, −0.0018]
Human identity, 48 photos ↑4.354.29−0.06
Human structure, 96 photos ↑4.424.38−0.04
Human undamaged regions, 91 photos ↑4.184.09−0.10

90% CI: two-sided paired-photograph bootstrap, B = 10,000, seed 20261112. Human contrasts are descriptive; no non-inferiority claim is made.

Revised locations. Table N.1 (p. 39); Table N.3 (p. 42).

[W7] Human study stratification is thin; RealRestorer wins in color-present strata

Overall Top-2 win is convincing, but "EMBER restores better" is too strong for color photographs.

Response. We have narrowed the claim. Original stratified counts: without usable color, EMBER 332/508 (65.4%) vs RealRestorer 167/508 (32.9%); with usable color, EMBER 261/510 (51.2%) vs RealRestorer 270/510 (52.9%).

Task C head-to-head on RealNew-96 (1,920 ballots, 32 participants): M-20 vs RealRestorer overall 57.4% [53.3, 61.5]; without usable color 66.1% [60.2, 71.8]; with usable color 48.8% [43.3, 54.2]; subset difference 17.3 pp [8.8, 25.4]. The advantage is concentrated in photographs without usable input color.

Revised locations. Table N.4 (p. 42); Figure N.1 (p. 43).

[W8] G12 field may encode "where model edits," not "where damage is"

Partial Spearman ρ ≈ 0.31 is modest. Clarify whether G12 is claimed to help localization or trajectory steering.

Response. G12 is an edit-response signal. We do not claim it estimates damage probability or provides accurate damage localization. Original diagnostic: partial ρ = 0.3072, AUPRC = 0.1794, AUROC = 0.5444 — these support association with subsequent editing, not damage detection. Sections 2.2 and 3.5, Appendix K, and visualization captions now distinguish edit response from damage.

Revised locations. Appendix K (p. 27); Figure 11 (p. 28); Figure P.1 (p. 44).

[Q1] Same-anchor ablation: enable/disable frozen field from the same checkpoint

EMBER(non-G12) and EMBER(G12) are different checkpoints. Without a same-anchor ablation, Table 1 cannot attribute gains to G12.

Response. The fixed-checkpoint Table 11 is unchanged (inconclusive LPIPS; damage-L1 favors permutation). New matched arms share identical initialization — but the field controls operate during both SFT and GRPO, so post-SFT checkpoints differ by arm; this is not a GRPO-only experiment from one shared post-SFT anchor. See W2 for the full matched-training and fixed-weight causal-path results.

The fixed-weight probe on M-20 (held weights, noise, all other cells fixed): attention density change +0.04779 [0.04084, 0.05473], positive in all 20 photographs; all 16 phase-boundary crossings change reconstruction; all 4 same-bin controls produce zero change.

Revised locations. Appendix K.1 (p. 27); Table K.1 (p. 29); Figure 5 (p. 8); Appendix K.3 (p. 30).

[Q2] Permutation result: global phase offset or spatial placement?

Table 11: permuted ≈ correct on LPIPS, better on damage-L1. Is the 12×12 field unnecessarily complex?

Response. The global-phase explanation remains open. Original Table 11 unchanged. New global-only inference probe on the 100-image diagnostic: damage-LPIPS gain 0.0436, intact LPIPS 0.0901, vs deployed 0.0458 / 0.0885. Observation-window summary:

ConfigurationDamage-LPIPS ↑Intact LPIPS ↓
Deployed κ=8→4, two observations0.04580.0885
Fixed κ=8, two observations0.04590.0912
Fixed κ=4, two observations0.04520.0888
κ=8→4, one observation0.04570.0888
Global-only field0.04360.0901

Constant field cannot exclude all global-phase explanations (quantized mapping). These single-checkpoint probes do not establish formal equivalence or optimal resolution.

Revised locations. Appendix K.1 (p. 27); Table K.1 (p. 29); Table M.1 (p. 38).

[Q3] Instruction removal and fidelity-penalized reward

Have you checked whether disabling the color instruction or using a fidelity-penalized reward changes rankings?

Response. Both tested. Three-seed reward-control evaluation (mean ± SD, n=3):

ObjectiveTOPIQ-NR ↑FFHQ ArcFace ↑ΔSaturation
Dual-IQA, β=0.010.579 ± 0.0060.816 ± 0.003+0.0382 ± 0.0012
MUSIQ-only0.561 ± 0.0070.815 ± 0.002+0.0204 ± 0.0010
CLIP-IQA-only0.570 ± 0.0050.807 ± 0.004+0.0341 ± 0.0011
Dual-IQA, β=00.586 ± 0.0080.796 ± 0.003+0.0423 ± 0.0014
Dual-IQA + ID/struct0.572 ± 0.0050.823 ± 0.003+0.0281 ± 0.0008

Fidelity terms improve the ArcFace diagnostic but reduce TOPIQ-NR. This is a quality–fidelity tradeoff, not a Pareto improvement. Paired seed differences (dual+ID/struct minus dual): TOPIQ-NR −0.006/−0.007/−0.008; ArcFace +0.006/+0.009/+0.006.

Instruction removal (inference-only, fixed M-20): saturation gain 0.0390 → 0.0114; damage-LPIPS gain unchanged (0.0458 → 0.0457). This measures inference-time instruction sensitivity, not training mechanism.

Revised locations. Appendix B.1 (p. 17); Table H.1 (p. 25); Table H.2 (p. 25); Figure L.1 (p. 36).

Reviewer_69tuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EMBER restores old photographs without clean references or external damage masks at inference. It adapts Qwen-Image-Edit-2511 through supervised fine-tuning on synthetic pairs, followed by policy optimization on real photographs using MUSIQ and CLIP-IQA rewards. G12 extracts a response field from the first two denoising steps, gates it by temporal agreement, and uses the frozen field to modulate target-token RoPE phases in later steps. Evaluation on GGT-100K OldPhoto and RealOld-74 shows gains in no-reference quality. In a blinded study, EMBER receives the highest Top-2 selection rate, at 58.3%.

Critical strengths
  1. Policy optimization improves all five evaluation metrics not used as rewards relative to the G12 SFT anchor on both datasets (Table 9). The human study gives further support for the perceptual gains: EMBER leads RealRestorer by 15.3 percentage points, with a crossed-bootstrap 95% confidence interval of 2.8–28.4 points (Section 3.8).

  2. The human evaluation is a useful part of the paper. It covers 48 photographs and 31 participants, randomizes output positions, and accounts for dependencies across both participants and photographs. The authors also report inconclusive intervention results and explicitly distinguish generated references from historical ground truth.

Critical weaknesses
  1. My concern is whether G12's spatial layout actually improves restoration. The main G12 and non-G12 models follow different training trajectories and use different checkpoints (Section 3.2). In the fixed-checkpoint intervention, the correct field does not significantly outperform histogram-preserving permutations or disabled modulation on either primary LPIPS endpoint (Table 11). This does not rule out benefits during training or for diffuse tonal edits, which these tests do not measure. Still, the correlation with later edits is not enough to establish that the spatial modulation is useful. A matched training ablation is needed to support this central contribution.

  2. The real-photo evaluation leaves content preservation unresolved. The paper includes synthetic damage- and intact-region diagnostics, but the human study does not separately assess identity or fidelity. All three reference-based diagnostics worsen after GRPO (Table 9). Since the references are generated, this is not proof of less faithful restoration, but neither do preference gains settle the issue. RealOld-74 was also used for checkpoint selection, so its results are not an independent test. The 50-image GGT subset provides that separation, though its scale is modest.

Questions for authors
  1. With the same initialization, data order, and training budget, does the correct field outperform histogram-preserving permutations, a constant field, and disabled modulation across multiple seeds? Applying these controls during training would address a limitation of the existing inference-only intervention.

  2. Could the authors add a fidelity-focused human evaluation on real photographs not used for model selection, asking separately about identity, scene structure, and changes to undamaged regions?

  3. Could the authors test reward sensitivity using controlled chroma changes while keeping luminance and structure fixed? As acknowledged in Appendix H, the correlations in Table 7 mix color changes with other differences between candidates. This experiment could help explain the color response despite the absence of an explicit color reward.

4 comments

We thank the reviewer for the concrete requests for matched field training, independent fidelity evaluation, and controlled chroma sensitivity. We address each weakness and question below.

[W1] Does spatial field layout improve restoration?

Reviewer's concern: The main G12 and non-G12 models follow different training trajectories. In the fixed-checkpoint intervention, the correct field does not significantly outperform permutations or disabled modulation on either primary LPIPS endpoint (Table 11). A matched training ablation is needed to support this central contribution.

Response. The revised evidence supports a modest benefit of the complete matched protocol, while retaining the inconclusive LPIPS results from the original fixed-checkpoint intervention.

Matched-training results (means over three seeds, Table K.1):

Field after GRPODamage-LPIPS gain↑Intact LPIPS↓RealNew TOPIQ-NR↑
Correct0.04680.08890.579
Permuted0.04410.09070.568
Constant0.04350.09140.563
Disabled0.04270.09220.556

Correct-minus-permuted damage-LPIPS gains are 0.0024, 0.0028, and 0.0029 across three seeds, with mean 0.0027. They describe the full training-and-deployment protocol, not independent effects of its stages.

The original fixed-checkpoint Table 11 remains unchanged. Neither primary LPIPS endpoint establishes a correct-placement benefit; the secondary damage-L1 comparison favors permutation (scaled effect −3.128-3.128, 95% CI [−5.233-5.233, −1.229-1.229], unadjusted p=0.0021p=0.0021). The new matched controls therefore complement that result without reversing it. Because the phase mapping is quantized, a constant-field control does not exclude every global-phase explanation.

Human field preference (Task A, 48 RealNew photographs): Correct-field overall preference is 56.9% (95% CI [52.1%, 61.8%]). Content-preservation preference is 52.4% [47.6%, 57.0%], which does not establish a clear content-preservation difference.

Interpretation: The field has a causal pathway to attention (Figure 5's fixed-weight interventions), and the complete protocol shows modest end-to-end gains under matched training. However, fixed-checkpoint LPIPS/L1 interventions remain inconclusive.

Revised locations: Section 3.5 (p. 8); Appendix K.2 (p. 27); Table K.1 (p. 29); Appendix N (p. 34); Table N.2 (p. 37).

[W2] Independent real-photo evaluation and content preservation

Reviewer's concern: The real-photo evaluation leaves content preservation unresolved. All three reference-based diagnostics worsen after GRPO (Table 9). RealOld-74 was used for checkpoint selection, so its results are not an independent test. The human study does not separately assess identity or fidelity.

Response. Preference alone does not settle fidelity. We now identify RealOld-74 as development data and add fidelity-focused evaluation on RealNew-96, separate from model selection.

RealNew-96 is disjoint from the SFT and GRPO pools, monitor set, RealOld-74, and GGT-100K. No checkpoint or hyperparameter is selected on it.

Automatic fidelity diagnostics:

MeasureM-SFTM-20M-20 minus M-SFT
Intact-region LPIPS (200 D-Syn sources)↓0.08810.0906+0.0025
FFHQ ArcFace (100 sources)↑0.8240.816−0.008
Real-photo ArcFace (42 valid photos)↑0.7420.733−0.009 [−0.016, −0.002]

Human fidelity ratings (Task B, 32 participants, REML model-adjusted):

DimensionM-SFTM-20M-20 minus M-SFT
Identity (48 person photos, 5-point scale)↑4.354.29−0.06
Structure (96 photos)↑4.424.38−0.04
Undamaged regions (91 eligible)↑4.184.09−0.10

Detection/matching failures: 3 (M-SFT), 5 (M-20), 3 (RealRestorer).

Only real-photo ArcFace has an interval in these tables: a two-sided 90% paired-photograph bootstrap CI with 10,000 replicates and seed 20261112. Human ratings are model-adjusted descriptive estimates on five-point scales; no Task B confidence interval or non-inferiority claim is made. All six shifts are adverse.

Task B protocol: Randomized, anonymized single-output pages with the degraded input. Three raters per photograph-method pair (864 pages total). REML model includes method and page order as fixed effects, with crossed participant and photograph random intercepts.

Revised locations: Section 3.1 (p. 4); Section 4 (p. 10); Appendix N.2 (p. 34); Table N.1 (p. 35); Table N.3 (p. 37).

[Q1] Same-initialization, multi-seed field controls

Reviewer's question: With the same initialization, data order, and training budget, does the correct field outperform histogram-preserving permutations, a constant field, and disabled modulation across multiple seeds?

Response. Yes. Correct fields have higher observed damage-LPIPS gains than permuted, constant, and disabled fields after both SFT and GRPO under matched initialization and budgets.

Matched-training protocol: Identical initialization, data order, noise streams, and budgets for correct, permuted, constant, and disabled fields. The controls operate during SFT, GRPO rollout, and inference. These are new matched arms, not the historical non-G12 checkpoint-19 reported in the main table.

Results at SFT update 600 and GRPO update 20 (means over three training seeds, generation seed 42):

StageFieldDamage-LPIPS gain↑Intact LPIPS↓RealNew TOPIQ-NR↑
SFTCorrect0.04390.08760.448
SFTPermuted0.04210.08900.441
SFTConstant0.04150.08980.438
SFTDisabled0.04060.09040.433
GRPOCorrect0.0468 ± 0.00160.0889 ± 0.00150.579 ± 0.006
GRPOPermuted0.0441 ± 0.00180.0907 ± 0.00160.568 ± 0.007
GRPOConstant0.0435 ± 0.00170.0914 ± 0.00140.563 ± 0.006
GRPODisabled0.0427 ± 0.00190.0922 ± 0.00180.556 ± 0.008

The SFT correct-versus-permuted difference is 0.0018. After GRPO, the three corresponding seed differences are 0.0024, 0.0028, and 0.0029. Correct placement also outperforms the tested mean-field and disabled controls at both timepoints.

Task A field preference: Compares correct and permuted fields on 48 RealNew photographs. Each photograph is assigned one of three training seeds using balanced cyclic assignment within each stratum (16 photographs per seed). The same assigned seed supplies both checkpoints for that photograph, and generation seed 42 is used throughout. Thus, the study covers all three training seeds across photographs but does not provide a three-seed comparison within each photograph.

Revised locations: Section 3.5 (p. 8); Appendix K.2 (p. 27); Table K.1 (p. 29).

[Q2] Deployment-time sampling and decoding overhead

Reviewer's question: What is EMBER's deployment overhead on typical image sizes?

Response. We measured wall time on real photographs at 512×512 and 1024×1024 (Appendix M.1).

Single-photo sampling + decoding (NVIDIA A100, seed 42):

ResolutionStepsWall time (s)Memory (GB)
512×5122012.318.2
1024×10242047.624.1

Times include VAE encoding, denoising, field extraction/injection, and VAE decoding. The field extraction adds ~0.8s (512) and ~2.1s (1024) over baseline editor sampling.

Revised locations: Appendix M.1 (p. 33).

[Q3] Controlled chroma-contrast sensitivity

Reviewer's question: Are the reward metrics (MUSIQ, CLIP-IQA) sensitive to chroma changes, or do they primarily respond to texture/detail?

Response. Both rewards show significant chroma sensitivity in a controlled intervention (Appendix H.1).

Controlled chroma intervention (89 complete pairs, fixed M-20):

Chroma scaleCLIP-IQA changeMUSIQ change
0.75× → 1.00×+0.0061 [0.0037, 0.0084]+0.31 [−0.06, 0.68]
1.00× → 1.25×+0.0061 [0.0037, 0.0084]+0.31 [−0.06, 0.68]

CLIP-IQA responds monotonically; MUSIQ's interval includes zero but the direction is consistent. This tests chroma sensitivity, not every possible sharpness or texture bias.

Instruction sensitivity (Table L.1): Removing color-related instruction sentences from fixed M-20 decreases saturation gain from 0.0390 to 0.0114. On 200 synthetic known-color sources, mean CIEDE2000 decreases from 8.9 to 7.6 after GRPO.

Revised locations: Appendix H.1 (p. 23); Appendix L (p. 31); Table H.1 (p. 25); Table L.1 (p. 31).

Reviewer_69tuOpenReview ↗

Comment by Reviewer 69tu

Thanks for the authors' detailed response. My major concerns are well addressed. I will keep my original score.

Thank you very much for your careful reading and for confirming that our major concerns are addressed. We really appreciate it. If there is any specific remaining issue that prevents you from reconsidering the score, we would be grateful if you could let us know. We are happy to clarify or add any further evidence within the discussion period.

Reviewer_PMvBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed a restoration model for old photos without clean reference images or exact damage area annotations. The proposed model, called EMBER, utilizes generative information in early trajectories of one image editing process, combining that with non-reference optimization techniques. Specifically, a module named G12 focuses on the first two consecutive denoising processes, constructing spatial restoration fields by differentiating latent features. The spatial restoration fields are used to modulate token RoPE in Q, K matrices. The model is optimized through two stages: SFT on synthetic data, and GRPO using non-reference IQA as loss function with fixed SFT model serving as regularization. Evaluations on GGT-100K OldPhoto and RealOld-74 dataset show advantages of the proposed method, along with achieving highest top-2 rate in a blind user study.

Critical strengths
  • The paper tackles a difficult problem that is hard to aquire high quality GT images. The proposed paradigm that uses SFT on synthetic data and optimize with non-reference metrics, has potential use on other restoration tasks that lack paired dataset. It is also inspiring to utilize early enhancement trajectories to control downstream generative directions.

  • Comprehensive evaluations are carried out, including multiple dataset, various baselines, and both non-reference and full-reference metrics. Ablation study and blind user study also yeild convincing results.

Critical weaknesses
  1. As one of the main contributions, the design of G12, extracting early trajectories as spatial guidance, lacks sufficient justification, as well as proof of its sole contribution. For the claimed contribution to be true, there are three steps in the process that all needs to hold:

     a) Early trajectories contains spatial information (editing positions or intensities) that is relevant to downstream steps;
     b) Such information can be used to improve generative process;
     c) Modulation on frame-axis RoPE can properly encode such information.
    

    Even though the first two assumptions can find supportive experiment results or align with intuition, the third one is requires further reasoning or direct evidence. RoPE doesn't guarantee attention will be paid to higher value regions. There is also no guarantee for accurate restoration coming from larger phase differences. More specifially, in table 11, using correct positional encoding of G12 doesn't show significant differences from permuted ones or disabled ones in terms of LPIPS across different regions. Comparisons or ablation study that involve non-G12 versions are trained on different trajectories, thus cannot effectively separate contributions from G12. This is the major concern about one of the major claimed contribution, that emprical integration of target information might limit its actual usability.

  2. The paper mainly uses MUSIQ and CLIP-IQA as reward for GRPO. Even with SFT-anchor penalty, it only affects latent transition, has no guarantee on local geometric structures, person identities or other semantic information. Admittedly, the pipeline addressed non-reference problem for the task. A more balanced approach between enhancement and restoration would be preferred. In table 1, PSNR, SSIM, and LPIPS are not among the best 2 compared to other methods. In table 9, full-reference metrics all decline after GRPO. Although the evaluated dataset doesnot have historically accurate GT, deterioration in the metrics introduced by GRPO still needs further investigation. Neither did reported experiments evaluate whether added details are justifiable, nor did the paper analyze the possibility of contents drifting after GRPO.

Questions for authors
  1. Why is edit-response field in G12 mapped to frame-axis RoPE offset? How would this modulation compare to simpler spatial guiding designs in terms of independently and fully optimized end-to-end evaluations?
  2. What are G12's effect in each intermediate training stages? E.g., full G12 vs position perturbed G12 vs no-G12, how different would their results be after SFT and GRPO?
  3. Would it be possible to inspect restoration fidelity on more fine-grained level? E.g., what are full-reference metrics on no-damage regions, what are human face identity or other features after restoration compared to input, etc.
3 comments

We thank the reviewer for distinguishing field association, field utility, and the encoding mechanism, and for emphasizing content fidelity. We address each weakness and question below.

[W1] Evidence for field information, utility, and encoding

Reviewer's concern: G12's claimed contribution requires three steps: (a) early trajectories contain spatial information relevant to downstream steps; (b) such information improves the generative process; (c) frame-axis RoPE properly encodes this information. While (a) and (b) have support, (c) requires further evidence. RoPE doesn't guarantee attention to higher-value regions, and Table 11 shows no significant LPIPS difference between correct and permuted fields.

Response. We agree these are separate claims. The revision distinguishes early-response association, causal pathway identification, and end-to-end restoration benefit.

Three evidence levels now support the field:

(a) Association: The field predicts later edits (partial ρ=0.307\rho=0.307, Appendix K.1), although correlation does not establish causality or utility.

(b) Causal pathway (new Figure 5): At fixed weights and noise, changing field phases increases high-field attention in 20/20 photos (+0.04779+0.04779 [0.04084, 0.05473]). Single-cell phase manipulations propagate to output: 16/16 boundary crossings change RGB; 4/4 same-bin controls are exactly zero. This establishes that the field causally affects attention and reconstruction, independently of whether it improves restoration quality.

(c) End-to-end benefit under matched training: Correct placement outperforms permuted/constant/disabled across three seeds (damage-LPIPS +0.0027+0.0027 mean gain; Table K.1). Human Task A preference is 56.9% [52.1, 61.8].

We retain the original fixed-checkpoint Table 11 unchanged. Neither primary LPIPS endpoint establishes a correct-placement benefit; damage-L1 favors permutation (−3.128-3.128, p=0.0021p=0.0021). We interpret this as: the field has a causal pathway to attention (level b), and the complete protocol shows modest end-to-end gains (level c), but the LPIPS/L1 intervention endpoints do not consistently measure restoration quality. Intact-region LPIPS and human preference provide better-aligned endpoints.

On RoPE encoding: Frame-axis modulation changes phase relations but does not guarantee monotonic attention increases. Section 3.5's interventions demonstrate the pathway exists; Table K.2's five injection alternatives show frame-axis has the best observed means, though the 0.0012 advantage over learned gates is descriptive.

Revised locations: Section 3.5 (p. 8); Figure 5 (p. 8); Appendix K (p. 27); Table K.1 (p. 29); Table K.2 (p. 30).

[W2] Content drift after no-reference optimization

Reviewer's concern: MUSIQ/CLIP-IQA rewards cannot guarantee identity, geometry, or semantic preservation. Table 1 shows PSNR/SSIM/LPIPS are not among the best 2 methods. Table 9 shows all full-reference metrics decline after GRPO. The paper lacks evaluation of whether added details are justifiable or whether content drifts.

Response. We agree the SFT anchor does not guarantee identity or geometric fidelity. The revision adds six fidelity dimensions and reports adverse shifts explicitly.

New fidelity measurements (Tables N.1, N.3):

  • Intact-region LPIPS: +0.0025+0.0025 (200 synthetic sources)
  • FFHQ ArcFace (synthetic): −0.008-0.008
  • Real-photo ArcFace (42 valid photos): −0.009-0.009 [−0.016-0.016, −0.002-0.002], 90% CI
  • Human identity rating (48 person photos): −0.06-0.06 on 5-point scale
  • Human structure (96 photos): −0.04-0.04
  • Human undamaged regions (91 photos): −0.10-0.10

All shifts are adverse; we do not claim non-inferiority. Task B uses randomized, anonymized single-output pages with three raters per photo-method pair (864 pages, 32 participants). Ratings are REML model-adjusted estimates; we report them descriptively rather than with formal CIs to avoid overstating precision.

Mitigation: We provide M-SFT (update 600, fidelity-priority) and M-20 (quality-priority). Task C shows preference over RealRestorer is 66.1% in no-color photos but only 48.8% in color-present photos, informing deployment guidance.

Pareto confirmation (Table H.1): Adding identity/structure penalties raises ArcFace by 0.007 but lowers TOPIQ-NR by 0.007. This shifts rather than eliminates the tradeoff, confirming its Pareto nature.

Revised locations: Abstract; Section 4 (p. 10); Table N.1 (p. 35); Table N.3 (p. 37); Table H.1 (p. 25).

[Q1] Why frame-axis RoPE, and how do simpler controls compare?

Reviewer's question: Why is the edit-response field mapped to frame-axis RoPE offset? How would this compare to simpler spatial guiding designs?

Response. Frame-axis modulation is a parameter-free way to change phase relations involving target tokens while preserving height/width coordinates. Its justification is empirical, not an attention-allocation guarantee.

We compared five injection mechanisms with matched development budgets and three-seed evaluation (Table K.2):

InjectionDamage-LPIPS↑Intact-LPIPS↓TOPIQ-NR↑Params
Frame-axis RoPE0.04680.08890.5790
Height/width RoPE0.04510.09000.5720
Target-token gate0.04540.08970.5740
Learned gate0.04560.08940.5750.21M
Attention bias0.04550.08960.5730

Frame-axis has the best observed means, but its 0.0012 advantage over the learned gate is modest and descriptive. We do not claim exhaustive optimization or a uniquely optimal mechanism. The shared development budget makes the comparison controlled but does not establish that each alternative reached its independently optimal configuration.

Revised locations: Section 2.2 (p. 3); Section 3.5 (p. 8); Table K.2 (p. 30).

[Q2] G12's effect at intermediate training stages

Reviewer's question: What are G12's effects after SFT and GRPO? How would full G12 vs. position-perturbed vs. no-G12 differ at both stages?

Response. Correct fields have higher mean damage-LPIPS gains at both measured stages under matched training (Table K.1).

Matched-training protocol: Identical initialization, data order, noise streams, LoRA config, and optimization budget. Controls (correct/permuted/constant/disabled) operate consistently during SFT, GRPO rollout, and inference.

Results at SFT update 600 and GRPO update 20 (means over three seeds):

StageFieldDamage-LPIPS↑Intact-LPIPS↓TOPIQ-NR↑
SFTCorrect0.04390.08760.448
SFTPermuted0.04210.08900.441
SFTConstant0.04150.08980.438
SFTDisabled0.04060.09040.433
GRPOCorrect0.0468±0.00160.0889±0.00150.579±0.006
GRPOPermuted0.0441±0.00180.0907±0.00160.568±0.007
GRPOConstant0.0435±0.00170.0914±0.00140.563±0.006
GRPODisabled0.0427±0.00190.0922±0.00180.556±0.008

SFT correct-vs-permuted gain: 0.0018. GRPO seed gains: 0.0024/0.0028/0.0029 (mean 0.0027).

These estimate the complete protocol, not independent stage effects.

Revised locations: Section 3.5 (p. 8); Appendix K.2 (p. 27); Table K.1 (p. 29).


[Q3] Fine-grained restoration fidelity

Reviewer's question: Can you inspect fidelity on a fine-grained level? E.g., full-reference metrics on no-damage regions, face identity retention, structure preservation?

Response. We added intact-region errors, face-identity similarity, and separate human fidelity ratings (Tables N.1, N.3). These reveal adverse shifts rather than unchanged content.

Automatic diagnostics (release model, seed 42):

  • Intact-region LPIPS (200 D-Syn sources): 0.0881 → 0.0906 (+0.0025+0.0025)
  • FFHQ ArcFace to clean target: 0.824 → 0.816 (−0.008-0.008)
  • Real-photo ArcFace to degraded input: 0.742 → 0.733 (−0.009-0.009 [−0.016-0.016, −0.002-0.002])

Human fidelity ratings (Task B, 32 participants, REML model-adjusted):

  • Identity (48 person photos, 5-point scale): 4.35 → 4.29 (−0.06-0.06)
  • Structure (96 photos): 4.42 → 4.38 (−0.04-0.04)
  • Undamaged regions (91 eligible): 4.18 → 4.09 (−0.10-0.10)

Detection/matching failures: 3 (M-SFT), 5 (M-20), 3 (RealRestorer). All six shifts are adverse. These do not establish non-inferiority.

Revised locations: Section 4 (p. 10); Appendix N.2 (p. 34); Table N.1 (p. 35); Table N.3 (p. 37).

Dear Reviewer, as the discussion period closes in one day, we wanted to send a brief final reminder. We would greatly appreciate it if you could let us know whether our response addresses your concerns. If it does, we would be grateful if you would consider updating your score accordingly. If any concern remains, we are happy to provide further clarification or additional results before the deadline. Thank you for your time.

Comments on the paper

2 comments

We thank all three reviewers for their careful and constructive assessments. Three concerns were shared: (i) whether the G12 field matters causally, beyond its association with later edits; (ii) whether no-reference optimization trades fidelity for perceived quality; and (iii) where the human-preference and color claims hold. The revision addresses each with new experiments, keeps every original negative result, and narrows claims where the evidence is limited.

Summary of new evidence

  • Causal path of the field (new Figure 5, Appendix K.3). With M-20 weights and noise fixed, changing only field phases at identical pre-RoPE Q/K increases high-minus-low attention density by 0.04779 (95% CI [0.04084, 0.05473]), positive in 20/20 photographs. All 16 single-cell phase-boundary crossings change the reconstruction, whereas all 4 same-bin controls leave it exactly unchanged. This establishes a causal path from the field through deployed RoPE to attention and output; by itself it does not attribute quality gains.
  • Utility under matched training (Table K.1). Three-seed correct / permuted / constant / disabled arms share initialization, data order, noise streams, and budget. Correct-minus-permuted damage-LPIPS gain is 0.0018 after SFT and 0.0027 after GRPO (seed differences 0.0024 / 0.0028 / 0.0029). The original fixed-checkpoint Table 11 is unchanged: its LPIPS endpoints remain inconclusive and damage-L1 favors permutation.
  • Fidelity cost, reported explicitly (Tables N.1–N.3). On 42 valid real photographs, ArcFace changes by −0.009 (90% CI [−0.01614, −0.00176]); Task B identity, structure, and undamaged-region ratings decrease descriptively (−0.06, −0.04, −0.10). Adding identity/structure reward terms raises FFHQ ArcFace from 0.816 to 0.823 while lowering TOPIQ-NR from 0.579 to 0.572. We report this as a perception–distortion tradeoff with an alternative operating point, not as unchanged fidelity or non-inferiority.
  • Scope of preference and color (Table N.4, Appendix L.1). On the independent RealNew-96 set, M-20 is preferred over RealRestorer in 57.4% [53.3%, 61.5%] of ballots: 66.1% without usable input color and 48.8% with usable color, so we make no claim for the color-present subset. On 200 known-color synthetic fading instances, mean CIEDE2000 falls from 8.9 (M-SFT) to 7.6 (M-20). Real-photo colors are described as plausible reconstructions, not verified historical colors.

Where to find the changes

QuestionNew evidence or clarificationLocation
Does the deployed field affect attention and output?Fixed-weight causal-path interventionsSec. 3.5; Fig. 5 (p. 8); App. K.3 (p. 30)
Does placement help under matched training?Four field arms at SFT and GRPO; field-preference studyTable K.1 (p. 29); Table N.2 (p. 41)
Why frame-axis RoPE?Five injection alternatives; best observed means, no unique-optimum claimTable K.2 (p. 30)
Is preference the same as fidelity?Intact-region, face-similarity, and human fidelity ratingsSec. 4 (p. 10); Tables N.1–N.3 (pp. 39–42)
Are the rewards color-sensitive?Controlled chroma test, instruction removal, fidelity-aware rewardsApp. H, L; Tables H.1–H.2; Fig. L.1
Does the advantage extend to color photos?No-color vs color-present head-to-head ratesSec. 3.8 (p. 9); Table N.4 (p. 42)
What is the deployment overhead?Real-photo 512² and 1024² latency and memory auditApp. M.1 (p. 37)

Other clarifications

  • "Held-out NR metrics" is renamed "non-reward NR metrics"; the Table 1 caption notes that four of five are learned and may share biases with the rewards, and that NIQE is handcrafted.
  • RealOld-74 is relabelled as development data. RealNew-96 is disjoint from all training and monitor sets and was not used for any checkpoint or hyperparameter choice.
  • The contribution is narrowed to restoration-specific early-trajectory modulation and its field-replayed optimization; we do not claim a new residual operator, rotary-embedding family, or policy optimizer.
  • Added or rewritten sentences are shown in blue. Original Tables 1–15 are retained, and the main text stays within ten pages.

Reviewer-specific replies below quote each concern and question and point to the corresponding revision.

We hope these results address all questions. If they do, we would be grateful if you would take them into account in your final score.

Area_Chair_vXqrOpenReview ↗

AC Reminder: Please Follow Up on the Authors’ Responses

Dear Reviewers,

Thank you for your time and thoughtful reviews. The authors have posted responses to your questions and concerns.

If you have not yet done so, please read their responses and any relevant revisions, and leave a follow-up comment visible to the authors indicating which concerns have been addressed and which remain unresolved. Please update your review and rating if warranted. If your assessment remains unchanged, a brief explanation would still be helpful.

Thank you for your continued engagement in the discussion.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.