acceptodds
Withdrawn from ICLR 2027

USIR: Unified Single-Step Image Restoration via Reward-Guided Privileged Self-Distillation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image restoration is challenging due to diverse and coupled degradations. Existing all-in-one methods often rely on explicit degradation understanding or task-specific guidance, and their performance declines under complex degradations, while recent agentic systems further introduce iterative assessment, planning, and tool execution, incurring substantial computational overhead and potential error accumulation across stages. To this end, we propose USIR, a Unified Single-step Image Restoration framework that directly restores degraded inputs in a single forward pass. At its core is Reward-Guided Privileged Self-Distillation (RPSD), a post-training framework that combines self-distillation under asymmetric information with reward-guided policy improvement. Specifically, reward-guided self-distillation progressively improves a degradation-aware privileged policy through reward-guided positive and negative targets, while privileged self-distillation transfers its preferred restoration behavior into a blind generator conditioned on the low-quality input. In this way, degradation-aware restoration knowledge and reward preference are progressively internalized into the single-step deployment model. Extensive experiments demonstrate that USIR outperforms both all-in-one and agentic restoration methods in perceptual quality and fidelity, while achieving significantly higher inference efficiency than agentic methods. Code and model weights are available at https://anonymous.4open.science/r/anonymous-744F.

Primary area:
generative models
Keywords:
Image Restoration, Diffusion Models, Self-Distillation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ydXPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies all-in-one image restoration under mixed degradations while retaining a single forward-pass deployment model. It proposes Reward-Guided Privileged Self-Distillation, which improves a degradation-aware privileged policy using reward-derived positive and negative targets, then distills its behavior into an LQ-only blind policy. Experiments on MiO100 and Real-1000 compare USIR with all-in-one and agentic restoration methods.

Critical strengths

The central training formulation is coherent and usefully exploits the asymmetric information available in paired restoration training. In Section 3.3, the privileged policy uses degradation prompts and difficulty-aware conditioning, while the deployed blind policy is trained under a fixed condition; Equations 5-10 make the reuse of the same candidate group for both reward-guided improvement and teacher-to-student distillation concrete. This is a practical way to turn expensive training-time degradation knowledge into a simple test-time pipeline.

Critical weaknesses

1: The broad state-of-the-art comparison is not yet sufficiently controlled to establish that the headline gains arise from USIR/RPSD rather than from the unusually strong foundation-model backbone and training setup. Section 4.1 builds on FLUX.2-Klein-Base-9B, applies a custom 61,490-pair training mixture across Stage 1 and Stage 2, and fine-tunes with rank-64 LoRA. Yet Section 4.2 retrains only DA-CLIP and FAPE-IR on the same training set; most other all-in-one and agentic results are inherited under their respective systems. The internal ablation in Table 5 establishes that RPSD improves the authors’ blind baseline, but it does not compare against a same-backbone, same-data single-step baseline trained with a conventional supervised or adversarial post-training objective, nor against the strongest recent blind unified restoration alternatives. Further, Tables 1-5 report point estimates only, without seed variation or confidence intervals, despite several claimed improvements being relatively close in some metrics. This does not invalidate the results, but it weakens the causal claim that RPSD is the reason for the full headline advantage. A convincing resolution would be a controlled FLUX-based no-RPSD baseline under the same data/compute budget, plus variance over independent training runs and comparisons to the most directly relevant unified blind methods.

2: The evaluation partially optimizes and then reports the same metrics, which makes the evidence for general restoration quality less independent than the paper suggests. The composite reward is instantiated using LPIPS, PSNR, SSIM, and CLIP-IQA+ in Equation 17, while the principal reported outcomes in Tables 1-2 include PSNR, SSIM, LPIPS, and CLIPIQA. Optimizing these metrics is a reasonable design choice, and the independent MANIQA/MUSIQ results are encouraging, but the very large improvements on reward-overlapping metrics should not alone be interpreted as fully independent evidence of perceptual superiority. Reporting a larger set of held-out quality measures or a more detailed human study protocol and uncertainty analysis would increase confidence that the method improves restoration rather than preferentially optimizing its reward suite.

Questions for authors

In Section 3.2, ptaskp_{task} is described as a task-specific prompt for a known degradation, but the mixed-degradation training samples in Section 4.1 and Appendix A contain one to five degradations. How is ptaskp_{task} formed for a mixed sample, for example as a concatenation of all degradation prompts, a single composite prompt, or a sampled primary degradation? This matters because the privileged teacher's available information is central to the claimed distillation mechanism.

8 comments
AuthorsOpenReview ↗

R3.W1-1: Pretrained Backbone Analysis

We thank the reviewer for the constructive feedback. We address each comment below and will incorporate the corresponding clarifications and additional experimental results into the revised manuscript.

1. Backbone

1) Large-scale generative priors are increasingly adopted in image restoration.

We argue that leveraging powerful pretrained generative models in low-level vision tasks is not an "unusual" practice, as described by the reviewer, but rather an increasingly established research direction. Recent models seen below demonstrate that large pretrained generative models are already widely explored in low-level vision.

Table R3-1. Representative Low-Level Vision Methods Leveraging Large-Scale Generative Priors.

ScopeMethodGenerative BackboneBackbone ParametersVenue
image restorationFAPE-IRFLUX.1-dev12BCVPR 2026
image restorationBeyond Sequence ToolsFLUX-Kontext≈12BCVPR 2026
Super-resolutionODTSRQwen-Image20BCVPR 2026
Super-resolutionTEASRQwen-Image-251220BECCV 2026
Super-resolutionDreamSRFLUX.1-dev12BCVPR 2026
Super-resolutionFluxSRFLUX.1-dev12BICML 2025
Super-resolutionResFlow-TunerFLUX.1-dev12BECCV 2026
Unified image restorationUSIR (Ours)FLUX.2-Klein-Base9B

2) Comparisons with methods using similarly powerful generative priors.

We extend our comparisons to BST on Real-1000 and MiO100, following the dataset groupings reported in BST and retaining the same six evaluation metrics as our paper. As shown in Tables R3-2 and R3-3, USIR outperforms BST on 17 out of 18 metrics across the three Real-1000 groups. On the challenging MiO100 Group C, USIR achieves improvements across all six metrics, including PSNR (20.80 vs. 19.85), LPIPS (0.2253 vs. 0.3392), and MUSIQ (69.17 vs. 58.78). USIR also shows consistently strong performance against FAPE-IR, despite using a smaller generative backbone. These comparisons demonstrate that USIR's performance advantage cannot be attributed simply to using a larger pretrained generative model.

Table R3-2. Comparison with strong generative restoration methods on Real-1000.

MethodBackbonePSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
FAPE-IR12B+7B21.570.74310.13900.36310.415866.47
BST12B+72B21.750.74650.21360.32060.424257.78
USIR (Ours)9B22.910.78660.10450.43130.553371.68

3) Controlled comparisons with matched backbones.

To further disentangle the contribution of the pretrained generative prior from that of RPSD, we e additionally conduct a backbone-matched comparison with FAPE-IR: we additionally evaluate USIR using the same FLUX.1-dev (12B) backbone as FAPE-IR. Due to time constraints, we limit Stage-1 initialization to only 10,000 steps on a single A100 GPU. However, USIR achieves substantially better perceptual quality across all four perceptual metrics, with slightly lower SSIM but comparable PSNR. This further supports the effectiveness of RPSD beyond backbone differences.

Together with the existing component ablations in Table 5, these controlled experiments are intended to further separate the contributions of pretrained generative priors, privileged supervision, and RPSD.

Table R3-3. Additional backbone-controlled comparisons on MiO100 Group C.

MethodGenerative BackboneExternal Understanding/Planning ModelPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
FAPE-IR*FLUX.1-dev (12B)Qwen2.5-VL-7B19.760.57670.42650.28250.399939.51
BSTFLUX-Kontext (≈12B)Qwen2.5-VL-72B19.850.57650.33920.32980.474258.78
USIR (Ours)FLUX.2-Klein-Base (9B)None20.800.61160.22530.42970.671569.17
USIRFLUX.1-dev (12B)None19.840.57560.23150.40170.591367.61
AuthorsOpenReview ↗

R3.W1-2: Retraining baselines and comparing conventional objectives

2. Additional Retraining of All-in-One Baselines

Following the evaluation protocol of prior work [1–3], we initially used published baseline results. To strengthen the comparison, we additionally retrained PromptIR and MiOIR, strong all-in-one methods on the same 61,490-pair training set.

Table R3-4. Comparisons of retraining of all-in-one baselines on MiO100.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
Group APromptIR*20.140.62530.45930.22320.320539.84
MiOIR*21.30.67060.39240.23330.364344.81
DA-CLIP*19.780.62840.30710.30670.45152.66
FAPE-IR*20.340.64650.35730.32550.481746.99
USIR (Ours)22.690.72790.14560.44460.677870.26
Group BPromptIR*20.620.64350.44060.22750.331242.17
MiOIR*21.560.68240.39010.24130.373445.68
DA-CLIP*19.730.65740.29010.30020.439453.48
FAPE-IR*19.860.63790.36600.32540.494047.79
USIR (Ours)22.740.74070.14580.44670.681870.36
Group CPromptIR*18.810.53290.60130.16900.254227.98
MiOIR*19.950.56810.55260.17590.295333.00
DA-CLIP*17.990.49350.49660.21010.330036.54
FAPE-IR*19.760.57670.42650.28250.399939.51
USIR (Ours)20.800.61160.22530.42970.671569.17

Under the same training-data setting, USIR outperforms all four retrained baselines across all six metrics in Groups A–C. Compared with the strongest retrained baseline for each metric, USIR improves PSNR by 1.39, 1.18, and 0.85 dB, respectively. These results show that USIR’s advantage persists after matching the task-specific training data, although backbone and training-procedure differences remain.

References

[1] Xiong et al., Beyond Sequential Tools: A Unified VLM Agent System for Photographic Post-Processing via Dynamic Multi-Expert Fusion, CVPR 2026.

[2] Jiang et al., Multi-Agent Image Restoration, IJCV 2026.

[3] Jia et al., TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration, ECCV 2026.

3. Same-Backbone, Same-Data Single-Step Baseline Trained with a Conventional Supervised and Adversarial Post-Training Objective

We add a conventional supervised fine-tuning baseline using the same backbone, training data, and single-step inference protocol, alongside the existing blind adversarial post-training baseline. RPSD outperforms both baselines across all six metrics. Compared with supervised fine-tuning, it improves PSNR by 2.23 dB, reduces LPIPS by 0.0206, and increases MUSIQ by 4.34. These results strengthen the evidence for benefits beyond conventional restoration training with the same backbone and data.

Table R3-5. Conventional supervised and adversarial post-training objective on MiO100 Group B.

Training StrategyPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
Conventional Supervised Fine-Tuning (1-step)20.510.69130.16640.37210.597966.02
Stage-1 Adversarial Post-Training (1-step)19.950.68680.18060.39730.598765.47
RPSD (Ours)22.740.74070.14580.44670.681870.36
AuthorsOpenReview ↗

R3.W1-3: Additional Comparisons with Recent Blind Restoration Methods

4. Additional Comparisons with Strongest Recent Blind Restoration Methods

We extend our comparisons to BST and PixRestore-L on MiO100 Group C and Real-1000. On MiO100 Group C, USIR outperforms all compared methods across all six metrics, improving PSNR over BST by 0.95 dB, reducing LPIPS from 0.3392 to 0.2253, and increasing MUSIQ from 58.78 to 69.17. On Real-1000, USIR achieves 22.91 dB PSNR and 0.1045 LPIPS, compared with the strongest baseline values of 21.75 dB and 0.1390, respectively. These additional comparisons demonstrate that USIR remains competitive against recent strong restoration methods across both benchmarks.

Table R3-6. Additional comparisons with recent restoration methods on MiO100 Group C.

MethodPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
FAPE-IR19.470.56270.51670.20430.333732.76
FAPE-IR*19.760.57670.42650.28250.399939.51
BST19.850.57650.33920.32980.474258.78
PixRestore-L18.430.48810.46990.18320.267332.89
USIR (Ours)20.800.61160.22530.42970.671569.17

Table R3-7. Additional comparisons on Real-1000 under different numbers of degradations.

MethodVenuePSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
FAPE-IRCVPR 2621.570.74310.13900.36310.415866.47
FAPE-IR*CVPR 2620.850.73350.13510.38770.453467.87
BSTCVPR 2621.750.74650.21360.32060.424257.78
PixRestore-LarXiv 2618.800.56440.35210.28220.272154.35
USIR (Ours)22.910.78660.10450.43130.553371.68
Reviewer_4yVYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents USIR, a unified single-step image restoration framework for complex mixed degradations, aiming to eliminate the reliance on explicit degradation understanding and multi-stage inference in existing all-in-one and agentic restoration approaches. Built on FLUX.2-Klein-Base-9B, USIR first initializes a degradation-aware privileged policy through a dual-noise-stream design and adversarial post-training. The authors then propose Reward-Guided Privileged Self-Distillation , which jointly improves the privileged policy through reward-guided self-distillation and transfers its restoration capabilities to an LQ-conditioned blind policy via reward-weighted privileged self-distillation. Both processes share candidate samples and reward evaluations to reduce training overhead. Experiments on MiO100 and Real-1000, together with training-strategy comparisons, component ablations, efficiency analyses, and human preference studies, demonstrate the effectiveness of the proposed framework.

Critical strengths
  1. USIR demonstrates substantial performance improvements under complex mixed degradations, achieving LPIPS scores of 0.1456, 0.1458, and 0.2253 on the three MiO100 groups, respectively, significantly outperforming the evaluated baselines. Results on Real-1000 further support its robustness across diverse degradation combinations. Moreover, its single-pass inference substantially reduces latency compared with iterative agentic restoration pipelines, demonstrating strong practical value.

  2. Table 5 shows that Privileged Self-Distillation improves PSNR from 20.18 to 21.91 dB, while the complete RPSD framework further increases it to 22.74 dB, demonstrating the complementary benefits of the two distillation processes. Table 3 further shows that sharing candidate samples and reward evaluations reduces Stage 2 training costs from approximately 21 to 14.85 GPU-hours while maintaining comparable restoration quality. These results provide meaningful evidence of the framework's effectiveness and computational efficiency.

Critical weaknesses
  1. Several components of USIR are closely related to existing techniques. The noise-hybrid dual-stream design in Stage 1 resembles ODTSR , the reward-guided construction of positive and negative targets is closely related to DiffusionOPSD , and the asymmetric teacher-student conditioning shares conceptual similarities with D-OPSD . Although RPSD integrates these mechanisms for unified image restoration and enables joint training through shared candidate samples, the current experiments primarily establish component-level improvements and computational savings over sequential training. They do not sufficiently isolate the benefits attributable to the proposed joint optimization mechanism from those arising from combining existing techniques, limiting the assessment of its methodological novelty.

  2. USIR builds upon a powerful 9B pretrained generative model with additional degradation-aware initialization and two-stage training, whereas most external baselines differ substantially in model capacity and pretraining resources. Although the authors retrain selected baselines and provide within-backbone ablations, the current evidence does not fully disentangle the contributions of the foundation model, privileged supervision, and RPSD itself. Furthermore, LPIPS, PSNR, SSIM, and CLIPIQA+ are used both in reward construction and in the main evaluation. While independent MANIQA, MUSIQ, and human preference results provide supporting evidence, the human study excludes several strong agentic baselines. Consequently, claims regarding the relative effectiveness of RPSD and the generalizability of its perceptual improvements require more careful qualification.

[1] One-Step Diffusion Transformer for Controllable Real-World Image Super-Resolution [2] On-Policy Self-Distillation in Diffusion Models [3] D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models

Questions for authors
  1. Could the authors provide controlled comparisons between RPSD and conventional privileged distillation, direct reward-guided fine-tuning, and a sample-sharing variant without joint policy updates, using the same backbone, initialization, and training budget? In particular, clarifying whether the improvements arise from the synergistic interaction between the two distillation processes, rather than their combination and sample reuse, would materially affect my assessment of the methodological contribution.

  2. Could the authors provide more independent evidence of RPSD's effectiveness, including evaluation metrics excluded from reward construction and human preference comparisons against stronger baselines such as TIR-Agent ? Reporting statistical uncertainty and clarifying the quality–compute trade-off under the complete training budget would further help establish the robustness and attribution of the reported improvements.

[1] TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration

7 comments
AuthorsOpenReview ↗

R2.W1-1: Novelty and the Contribution of Joint Optimization

We thank the reviewer for the constructive comments. Our responses are provided below, and the corresponding clarifications and additional experimental results will be incorporated into the revised manuscript. We agree that individual mechanisms in USIR are related to prior work, and we do not claim them as our contribution. We first clarify how USIR relates to them, then provide controlled experiments that isolate the effect of our joint optimization.

1. What we reuse and what we change

Table R2-1. Comparison of USIR against prior work.

Component in USIRRelated WorkWhat we takeWhat is different in our setting
Noise-hybrid dual-stream (Stage 1)ODTSRThe idea of a second visual stream that receives the LQ latent with a separate noise level, while the main stream keeps the pretrained prior (already cited in Sec. 4.2).In ODTSR the control noise is an adjustable knob for fidelity and prompt controllability in real-world super-resolution. In USIR the condition-stream noise is set from the LQ–GT distance during training, so it encodes degradation severity as privileged information. It is removed at deployment (σcB=0\sigma_c^B = 0), and the model targets multiple, mixed degradations.
Positive/negative target constructionDiffusionOPSDBounded reward-gradient ascent/descent targets within a trust region (cited in Eq. 6).DiffusionOPSD aligns a single policy with rewards for text-to-image generation. We further adapt its original multi-step formulation to single-step generation, thereby eliminating the endpoint estimation error introduced by intermediate velocity predictions in multi-step optimization (theoretical justification in Reviewer RiQz W1.1). In USIR, this step improves a privileged generator, and the same candidate group is reused to distill a blind generator.
Asymmetric teacher–student conditioningD-OPSDThe idea that the same model can act as teacher and student under different conditioning contexts.In D-OPSD, the teacher context contains the target image, no external reward is used, and the goal is continually tuning the few-step distilled model with new concepts and styles. In USIR, the asymmetry is degradation information (task prompt and difficulty), the teacher is itself improved by rewards, and the goal is to internalize degradation understanding into a single blind generator for restoration.

Our contribution is therefore not any single mechanism, but:

  1. A new paradigm and a framework for complex image restoration: Internalizing degradation understanding into one blind generator through privileged-to-blind training, which removes the large MLLM at inference (e.g., Qwen2.5-VL-7B in FAPE-IR, Qwen2.5-VL-72B in BST, GPT-4o in AgenticIR). USIR needs only a single generator and one forward pass, at 0.78 s per image.
  2. Joint optimization over shared candidates. Reward-guided improvement of the privileged generator and privileged-to-blind distillation are performed in the same loop over the same candidates and rewards.
  3. Reward-weighted internalization, which makes the blind generator absorb preferred behaviors rather than the average behavior.
  4. Strong performance.

The experiments below isolate points 2 and 3.

AuthorsOpenReview ↗

R2.W1-2: Novelty and the Contribution of Joint Optimization

2. Controlled Comparisons

All methods use the same FLUX.2-Klein-Base-9B backbone and are evaluated under the same blind protocol on MiO100 Group B.

We factorize the design into {sequential, joint} × {uniform, reward-weighted} and add a variant that shares samples without joint updates.

Table R2-2. Controlled comparison of RPSD on MiO100 Group B.

ConfigurationPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑GPU-h ↓
Direct blind reward-guided fine-tuning20.560.69700.17200.39770.610667.3014.88
Conventional privileged self-distillation21.910.72570.15740.42870.671569.577.15
A. Sequential + uniform distillation21.480.72160.14720.44700.670870.3621.45
B. Sequential + reward-weighted distillation21.640.72450.14730.44720.671070.3821.70
Sequential (best order in Table 3: privileged SD → reward-guided SD)22.660.74010.14330.43940.676670.5121.16
Sample sharing w/o joint updates22.960.74970.16490.42630.659468.3222.85
C. Joint + uniform distillation22.530.73710.14860.43070.666669.6414.62
D. Joint + reward-weighted (RPSD)22.740.74070.14580.44670.681870.3614.85

Finding 1. Joint Optimization Improves Fidelity and Reduces Cost

With the same weighting, joint training improves PSNR/SSIM by about 1.1 dB / 0.016 (A → C: 21.48 → 22.53; B → D: 21.64 → 22.74) and reduces Stage-2 cost by about 30% (21.45–21.70 → 14.62–14.85 GPU-hours).

We attribute the fidelity gain to progressive learning: the blind generator tracks the improving privileged generator step by step, so each update faces a smaller target shift.

Finding 2. Reward Weighting Is Effective When Targets Are Paired with the Current Update

Under joint training, reward weighting improves all perceptual metrics (C → D: LPIPS 0.1486 → 0.1458, MANIQA 0.4307 → 0.4467, CLIPIQA 0.6666 → 0.6818, MUSIQ 69.64 → 70.36) together with PSNR (22.53 → 22.74).

Under sequential training, the same weighting brings almost no change (A → B: MUSIQ 70.36 → 70.38, CLIPIQA 0.6708 → 0.6710, LPIPS 0.1472 → 0.1473).

Without reward weighting, joint training is slightly lower than sequential on perceptual metrics (A → C: MUSIQ 70.36 → 69.64, CLIPIQA 0.6708 → 0.6666). Reward weighting recovers and surpasses this gap. Hence, the gains come from the combination of joint updates and reward weighting, not from either alone.

Finding 3. Sharing Samples Alone Is Not Sufficient

The variant that shares samples and reward evaluations but optimizes the two objectives separately achieves higher PSNR (22.96 vs. 22.74) and SSIM (0.7497 vs. 0.7407), but is worse on all four perceptual metrics: LPIPS (0.1649 vs. 0.1458), MANIQA (0.4263 vs. 0.4467), CLIPIQA (0.6594 vs. 0.6818), and MUSIQ (68.32 vs. 70.36).

We attribute this to privileged targets that are no longer paired with the current blind update, which encourages regression toward averaged results that favor pixel fidelity but suppress high-frequency detail.

Comparison with the Strongest Sequential Baseline

Against the best sequential order in Table 3 (privileged SD → reward-guided SD), RPSD achieves comparable quality at about 30% lower cost (21.16 → 14.85 GPU-hours). That baseline is slightly better on LPIPS (0.1433) and MUSIQ (70.51), and RPSD is better on other metrics. We therefore claim efficiency at comparable quality against the best sequential schedule, and larger gains against the other orders.

3. Summary

The controlled comparisons show that the benefit of RPSD is not explained by simply combining existing objectives or sharing samples:

  • Sample sharing alone favors distortion-oriented fidelity but degrades perceptual quality.
  • Joint updates improve fidelity and reduce cost by about 30%.
  • Reward weighting becomes effective under joint updates and gives the best fidelity–perception balance.

We will add Table R2-2, Fig. R1, and the discussion of ODTSR, DiffusionOPSD, and D-OPSD to the revised paper.

AuthorsOpenReview ↗

R2.W2: Isolating the Contribution of the Pretrained Backbone, Privileged Supervision, and RPSD

We thank the reviewer for the suggestions.

1. Backbone

1) Comparisons with methods using similarly powerful generative priors.

First, leveraging large-scale pretrained generative models is not unique to USIR. For instance, FAPE-IR employs FLUX.1-dev (12B) as its restoration backbone, together with an MLLM-based planner, whereas USIR builds on the smaller FLUX.2-Klein-Base-9B. Moreover, we include the recent Beyond Sequential Tools (BST, CVPR 2026), which also combines a pretrained diffusion backbone (FLUX-Kontext) with VLM-guided degradation understanding and expert adaptation.

2) Additional comparisons with BST.

We extend our comparisons to BST on Real-1000 and MiO100, following the dataset groupings reported in BST and retaining the same six evaluation metrics as our paper.

As shown in Tables R2-3 and R2-4, USIR outperforms BST on 17 out of 18 metrics across the three Real-1000 groups. On the challenging MiO100 Group C, USIR achieves improvements across all six metrics, including PSNR (20.80 vs. 19.85), LPIPS (0.2253 vs. 0.3392), and MUSIQ (69.17 vs. 58.78). USIR also shows consistently strong performance against FAPE-IR, despite using a smaller generative backbone. These comparisons demonstrate that USIR's performance advantage cannot be attributed simply to using a larger pretrained generative model.

Table R2-3. Comparison with strong generative restoration methods on Real-1000, grouped by the number of degradations following BST.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
Group 1FAPE-IR24.340.80840.09900.38670.436968.98
BST22.900.77180.17110.32950.409860.67
USIR (Ours)23.950.82180.09500.44520.569072.20
Group 2FAPE-IR19.790.69000.17890.34190.420763.83
BST21.100.71750.25280.29970.456955.84
USIR (Ours)22.420.76850.12130.41690.557170.25
Group 3FAPE-IR16.040.64400.17900.33300.323164.28
BST19.250.73100.26500.34500.388052.30
USIR (Ours)20.340.70460.09380.41910.484073.66

3) Controlled comparisons with matched backbones.

To further disentangle the contribution of the pretrained generative prior from that of RPSD, we additionally evaluate USIR using the same FLUX.1-dev (12B) backbone as FAPE-IR. Due to time constraints, we limit Stage-1 initialization to only 10,000 steps on a single A100 GPU. Nevertheless, USIR achieves substantially better perceptual quality across all four perceptual metrics, with slightly lower SSIM but comparable PSNR. This further supports the effectiveness of RPSD beyond backbone differences.

Together with the existing component ablations in Table 5, these controlled experiments are intended to further separate the contributions of pretrained generative priors, privileged supervision, and RPSD.

Table R2-4. Additional backbone-controlled comparisons on MiO100 Group C.

MethodGenerative BackboneExternal Understanding/Planning ModelPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
FAPE-IR*FLUX.1-dev (12B)Qwen2.5-VL-7B19.760.57670.42650.28250.399939.51
BSTFLUX-Kontext (≈12B)Qwen2.5-VL-72B19.850.57650.33920.32980.474258.78
USIR (Ours)FLUX.2-Klein-Base (9B)None20.800.61160.22530.42970.671569.17
USIRFLUX.1-dev (12B)None19.840.57560.23150.40170.591367.61

2. Controlled Comparisons

See our response to R2.W1 (Table R2-2) and R2.Q1 (Table R2-5).

3. Independent Evaluation and Reward Robustness

See our response to R2.Q2-1.

4. Human Preference Evaluation against Strong Baselines

See our response to R2.Q2-2.

5. Clarification of Claims and Attribution

We agree, and we qualify the claims as follows.

Regarding the relative effectiveness of RPSD, we attribute it only to the same-backbone controls (Tables 5, R2-2, and R2-4), and the paired-bootstrap confidence intervals in our response to R2.Q2 (Table R2-8) quantify the differences between RPSD and its components.

Regarding the generalizability of perceptual improvements, we point to the Real-1000 comparison in Table R2-3 (USIR is better than BST on 17 of 18 metrics, including the reward-independent MANIQA and MUSIQ) and to the human preference study (Table R2-7).

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

USIR is a single-step image restoration framework built on FLUX.2-Klein-Base-9B. It first trains a teacher with privileged degradation information, then combines reward-guided teacher updates with reward-weighted distillation into a blind student. Both updates share samples and reward evaluations to reduce training cost. Only the student is used at inference, without explicit degradation information or iterative tool execution. Experiments on MiO100 and Real-1000 report strong fidelity and perceptual quality, supported by ablations and a user study. Inference is faster than that of the evaluated agentic methods.

Critical strengths
  • Using degradation information during training and removing this requirement at inference is a reasonable design. The resulting single-step pipeline avoids the iterative calls used by agentic methods.
  • The reported restoration results are strong on MiO100 and Real-1000, with supporting visual comparisons and a user study. The same-backbone ablations also show gains from privileged distillation and the full framework.
  • Sharing teacher samples and reward evaluations makes the combined training more efficient. Table 3 shows lower training cost than sequential training while maintaining comparable restoration quality.
Critical weaknesses
  • The methodological novelty is limited. Reward-guided target construction and self-distillation largely follow DiffusionOPSD (Zhou et al., 2026). The main addition appears to be their combination with privileged distillation, whose new insight needs clearer explanation.
  • Key training comparisons are missing. Table 4 evaluates alternative optimizers only on the teacher, without comparing their final blind students. Direct reward-based blind post-training is also not evaluated.
  • External comparisons mix training-method gains with backbone and initialization differences. Same-data retraining is helpful, but comparisons with recent methods using similarly strong generative priors are needed to better assess RPSD's contribution.
Questions for authors
  • What changes does Reward-Guided Self-Distillation introduce beyond DiffusionOPSD?
  • How does the final blind student compare with direct reward-based blind post-training, and with students distilled from teachers trained using the alternative optimizers in Table 4?
  • Can the authors better isolate the backbone's contribution through comparisons with recent methods using comparable generative priors or the same FLUX Klein backbone?
4 comments
AuthorsOpenReview ↗

R1.W1: Motivation and Novelty

Thanks for the constructive comments. Responses to specific comments are given below, which are also be added to the revised paper.

Theoretical Advantage of Native Single-Step Generation

While our reward-guided optimization draws inspiration from DiffusionOPSD, we argue that native single-step generation provides a theoretically more consistent setting for reward-guided self-distillation by eliminating the discrepancy between the estimated clean endpoint and the actual generated output.

Proposition 1 (Endpoint Consistency). Consider a flow-based generator with an Euler sampling schedule 0=σ0<σ1<⋯<σK0=\sigma_0<\sigma_1<\cdots<\sigma_K. At an intermediate state zjz_j, DiffusionOPSD estimates the clean endpoint by extrapolating the current predicted velocity:

z^0=zj−σjvθ(zj,σj).\hat z_0=z_j-\sigma_j v_\theta(z_j,\sigma_j).

However, the actual endpoint obtained by completing the remaining denoising steps is

z0=zj−∑k=1jΔσkvθ(zk,σk),z_0=z_j-\sum_{k=1}^{j}\Delta\sigma_k v_\theta(z_k,\sigma_k),

where Δσk=σk−σk−1\Delta\sigma_k=\sigma_k-\sigma_{k-1}. Their discrepancy is therefore

e_j=\hat z_0-z_0 =\sum_{k=1}^{j}\Delta\sigma_k \left[v_\theta(z_k,\sigma_k)-v_\theta(z_j,\sigma_j)\right]. }$$ For the multi-step generation setting of DiffusionOPSD, this discrepancy is generally nonzero because the velocity field varies along the denoising trajectory. Assuming the velocity field is Lipschitz-continuous in both state and time, with constants $L_z$ and $L_\sigma$, and satisfies $\|v_\theta\|_2\leq M$, we obtain $$\boxed{ \lVert e_j \rVert_2 \leq \frac{L_z M + L_\sigma}{2} \left( \sigma_j^2 - \sum_{k=1}^{j} \Delta\sigma_k^2 \right) }$$ In contrast, for **native single-step generation** ($K=1$), the predicted endpoint exactly coincides with the actual generated output: $$\boxed{ \hat z_0=z_0,\qquad \|e_1\|_2=0. }$$ This equality holds regardless of the velocity prediction accuracy, since the predicted clean endpoint is itself the output of the single-step sampler. **Implications for USIR.** These results establish an important theoretical motivation for adopting native single-step generation in RPSD. Unlike multi-step approaches that may introduce an additional endpoint-estimation discrepancy, our formulation performs reward-guided optimization directly on the model's actual output. This eliminates the intermediate-to-terminal endpoint inconsistency, providing a more consistent optimization target for privileged-policy improvement and subsequent blind-policy distillation. ### Clarification of the Novelty, Motivation and Contribution of RPSD **1. Novelty.** We would like to clarify that the primary novelty of RPSD lies in its unified optimization framework, rather than the individual self-distillation techniques. To the best of our knowledge, USIR is the first work to introduce a reward-guided privileged self-distillation framework for unified image restoration, jointly improving a degradation-aware privileged policy and transferring its learned capabilities and reward preferences into a blind restoration model. **2. Motivation.** Our motivation stems from two complementary objectives in unified image restoration: 1. Transferring the behavior of a degradation-aware privileged policy to a blind policy that operates without privileged information at inference time. 2. Improving the restoration quality of a degradation-aware privileged policy through reward optimization, thereby preventing the blind policy from being constrained by the capabilities of the initial privileged policy. **3. Contributions.** Therefore, our contributions can be summarized as follows: 1. **A new paradigm and a framework for complex image restoration:** Internalizing degradation understanding into one blind generator through privileged-to-blind training, which removes the large MLLM at inference (e.g., Qwen2.5-VL-7B in FAPE-IR, Qwen2.5-VL-72B in BST, GPT-4o in AgenticIR). USIR needs only a single generator and one forward pass, at 0.78 s per image. 2. **Joint optimization over shared candidates:** Reward-guided improvement of the privileged generator and privileged-to-blind distillation are performed in the same loop over the same candidates and rewards. 3. **Reward-weighted internalization:** The blind generator absorbs preferred behaviors rather than the average behavior. 4. **Strong performance.**
AuthorsOpenReview ↗

R1.W2: Comparison of Final Blind Students under Different Post-Training Strategies

W2: Comparison of Final Blind Students under Different Post-Training Strategies

Direct reward-based blind post-training is applied Reward-Guided Self-Distillation (RGSD) to the blind-initialized model under a fixed prompt and fixed condition noise. As shown in Table R1-1, direct blind RGSD improves PSNR from 19.95 to 20.56 dB and reduces LPIPS from 0.1806 to 0.1720, demonstrating that reward-guided optimization is effective. However, incorporating privileged supervision through Privileged Self-Distillation further improves restoration performance, while the complete RPSD framework achieves 22.74 dB PSNR and 0.1458 LPIPS. These results highlight the additional benefits of transferring degradation-aware restoration capabilities and reward preferences into the blind policy, beyond direct blind reward optimization.

Regarding alternative teacher optimization strategies, we extend the comparison to the final blind students obtained by distilling teachers optimized with ReFL, Flow-GRPO, and DiffusionNFT, respectively. All variants use the same initialization, training data, and blind-student distillation protocol to isolate the influence of teacher optimization. The comparison is summarized in Table R1-1. RPSD achieves the best PSNR, SSIM, LPIPS, and MUSIQ, demonstrating that our unified optimization framework is more effective than alternative teacher optimization strategies.

Table R1-1. Comparison of blind restoration performance under different post-training strategies on MiO100 Group B.

ComparisonTraining StrategyPSNRSSIMLPIPSMANIQACLIPIQAMUSIQ
Initialization referenceBlind Initialization19.950.68680.18060.39730.598765.47
Direct blind post-trainingDirect Reward-Based Blind Post-Training20.560.69700.17200.39770.610667.30
Alternative teacher optimizers → blind studentsReFL → Privileged Distillation22.460.73530.14900.45110.689869.84
Flow-GRPO → Privileged Distillation21.960.72690.15810.43140.673669.69
DiffusionNFT → Privileged Distillation21.990.72890.15670.43080.674169.73
Our final blind studentRPSD (Ours)22.740.74070.14580.44670.681870.36
AuthorsOpenReview ↗

R1.W3: Backbone and initialization

W3: Backbone and Initialization

1. Comparisons with Methods Using Similarly Powerful Generative Priors

First, leveraging large-scale pretrained generative models is not unique to USIR. For instance, FAPE-IR employs FLUX.1-dev (12B) as its restoration backbone, together with an MLLM-based planner, whereas USIR builds on the smaller FLUX.2-Klein-Base-9B. Moreover, we include the recent Beyond Sequential Tools (BST, CVPR 2026) [1], which also combines a pretrained diffusion backbone (FLUX-Kontext) with VLM-guided degradation understanding and expert adaptation.

We extend our comparisons to BST on Real-1000 and MiO100 Group C, following the dataset groupings reported in BST and retaining the same six evaluation metrics as our paper. Since BST has not publicly released its code or model weights, we directly adopt the results reported in its original paper.

As shown in Tables R1-2 and R1-3, USIR outperforms BST on 17 out of 18 metrics across the three Real-1000 groups. On the challenging MiO100 Group C, USIR achieves improvements across all six metrics, including PSNR (20.80 vs. 19.85), LPIPS (0.2253 vs. 0.3392), and MUSIQ (69.17 vs. 58.78). USIR also shows consistently strong performance against FAPE-IR, despite using a smaller generative backbone. These comparisons demonstrate that USIR's performance advantage cannot be attributed simply to using a larger pretrained generative model.

Table R1-2. Comparison with strong generative restoration methods on Real-1000, grouped by the number of degradations following BST.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
Group 1FAPE-IR24.340.80840.09900.38670.436968.98
BST22.900.77180.17110.32950.409860.67
USIR (Ours)23.950.82180.09500.44520.569072.20
Group 2FAPE-IR19.790.69000.17890.34190.420763.83
BST21.100.71750.25280.29970.456955.84
USIR (Ours)22.420.76850.12130.41690.557170.25
Group 3FAPE-IR16.040.64400.17900.33300.323164.28
BST19.250.73100.26500.34500.388052.30
USIR (Ours)20.340.70460.09380.41910.484073.66

2. Controlled Comparisons with Matched Backbones

To further disentangle the contribution of the pretrained generative prior from that of RPSD, we additionally evaluate USIR using the same FLUX.1-dev (12B) backbone as FAPE-IR. Due to time constraints, we limit Stage-1 initialization to only 10,000 steps on a single A100 GPU. Nevertheless, USIR achieves substantially better perceptual quality across all four perceptual metrics, with slightly lower SSIM but comparable PSNR. This further supports the effectiveness of RPSD beyond backbone differences.

Together with the existing component ablations in Table 5, these controlled experiments further separate the contributions of pretrained generative priors, privileged supervision, and RPSD.

Table R1-3. Additional backbone-controlled comparisons on MiO100 Group C.

MethodGenerative BackboneExternal Understanding/Planning ModelPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIPIQA ↑MUSIQ ↑
FAPE-IR*FLUX.1-dev (12B)Qwen2.5-VL-7B19.760.57670.42650.28250.399939.51
BSTFLUX-Kontext (≈12B)Qwen2.5-VL-72B19.850.57650.33920.32980.474258.78
USIR (Ours)FLUX.2-Klein-Base (9B)None20.800.61160.22530.42970.671569.17
USIRFLUX.1-dev (12B)None19.840.57560.23150.40170.591367.61

[1] Xiong et al., Beyond Sequential Tools: A Unified VLM Agent System for Photographic Post-Processing via Dynamic Multi-Expert Fusion, CVPR 2026.

Dear Reviewer,

We hope this message finds you well.

As the discussion period is nearing its end with less than three days remaining, we wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we are eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

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

Related papers

Discussion (0)

Sign in to comment.