acceptodds
Under review as a conference paper at ICLR 2027

CURE: Clean-Guided Understanding Restoration for Degraded Images

Abstract

Image restoration aims to recover high-quality images from degraded observations, providing reliable inputs for downstream visual tasks. When these tasks are performed by large vision-language models (LVLMs), conventional restore-then-understand pipelines still optimize restoration primarily for pixel fidelity or perceptual quality, which does not directly guarantee improvements in understanding and may propagate restoration errors into model responses. To address this issue, we propose understanding restoration for degraded inputs and introduce CURE (Clean-Guided Understanding REstoration). CURE first lets the model generate responses to degraded images, then uses its predictive distributions under clean-image conditions to supervise each step of the generated responses. This supervision guides the model’s generation decisions toward those under clean conditions, aiming to restore visual understanding impaired by image degradation. During inference, the adapted model directly processes degraded images without an additional image restoration module. On Qwen3.5-4B, CURE improves the average score on the real-world degradation benchmark R-Bench from 63.49 to 67.38, outperforming the evaluated two-stage restoration pipelines. It also achieves overall improvements in cross-dataset evaluations with synthetic degradations and evaluations on clean images. Furthermore, CURE can complement image restoration models in a two-stage pipeline to improve understanding of degraded images.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visual Understanding, Robustness, Image Restoration, Large Vision-Language Models

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

For image restoration for large vision-language models (LVLM). This paper proposes CURE, a clean-guided understanding restoration framework for improving LVLM robustness under degraded images. Different from conventional restore-then-understand pipelines, CURE directly adapts LVLMs by aligning token-level prediction distributions between clean and degraded images using a clean-view teacher. Experiments on real and synthetic degradation benchmarks demonstrate consistent improvements over existing LVLMs and restoration-based approaches.

Critical strengths
  1. The paper identifies that pixel-level restoration quality does not necessarily correlate with downstream LVLM understanding, which is an important issue for multimodal robustness.

  2. The clean-guided distribution alignment objective is intuitive and technically reasonable.

  3. The paper evaluates multiple degradation settings and provides useful analysis of visual and language adaptation contributions.

Critical weaknesses
  1. The proposed objective is closely related to existing knowledge distillation and consistency learning methods. The paper needs to better clarify what is fundamentally different from standard degraded-image fine-tuning or self-distillation.

  2. The current ablation study does not sufficiently isolate the effect of the proposed clean-guided objective. It remains unclear whether the observed improvements are mainly attributed to the proposed prediction alignment mechanism or simply result from additional adaptation to degraded images.

  3. The term "understanding restoration" is ambiguous and requires further justification.

  4. Could the authors provide more efficiency analysis, such as inference latency, computational cost, memory consumption, and training overhead? Since CURE relies on full-parameter fine-tuning of the LVLM, additional discussion on its scalability to larger models would also be valuable.

Questions for authors
  1. How much of the performance improvement comes from the proposed clean-guided objective itself, compared with simply fine-tuning the LVLM on degraded images?

  2. Why is the proposed approach considered "understanding restoration" rather than general degraded-image robustness fine-tuning?

  3. What is the computational overhead of CURE compared with existing restore-then-understand pipelines?

  4. How scalable is CURE to larger LVLMs or parameter-efficient adaptation settings?

2 comments
AuthorsOpenReview ↗

Author Response – Part 1

Thank you for your valuable and constructive feedback.

Q1: Fundamental differences from standard degraded-image fine-tuning and self-distillation

A1: Thank you for this insightful comment. CURE is a clean-guided self-distillation method related to knowledge distillation and consistency learning. Unlike standard degraded-image fine-tuning that fits annotated labels and offline self-distillation that fits static pre-generated sequences, CURE uses the frozen model's clean-image branch as guidance (clean-guided) and directly aligns the model's next-token decision distributions along trajectories rolled out by the model on degraded images.

Specifically, this differs in two respects:

(1) Clean-guided full-distribution supervision.
Standard degraded-image fine-tuning relies on annotated hard labels, which can overfit to dataset-specific answer formats, while conventional knowledge distillation requires an external teacher. Instead, CURE requires neither annotations nor an external teacher, using the frozen model's clean-image output distribution to restore its clean-view decision preferences.

(2) Token-level decision alignment along the model's own trajectory.
Autoregressive generation is a sequence of dependent decisions. Offline self-distillation only fits static pre-generated responses. In CURE, prefixes are rolled out on the degraded image and token-level alignment is applied along these trajectories, correcting the model at the exact decision points it visits under degradation.


Q2: Ablation on the clean-guided objective and the prediction alignment mechanism

A2: Thank you for this valuable question, which prompted us to conduct more comprehensive experiments and analyses to strengthen our conclusions. We added two direct SFT baselines: GT-answer SFT and Clean-response SFT. The results show that both the clean-guided objective (+6.58 on average from GT-answer SFT to Clean-response SFT) and the prediction alignment mechanism (+1.63 on average from Clean-response SFT to CURE) contribute to CURE's gains, rather than merely resulting from additional adaptation to degraded images.

Both SFT baselines use the exact same training data size, number of training steps, and learning rate as CURE:

  • GT-answer SFT: takes degraded images as input, with GQA ground-truth answers as supervision targets;
  • Clean-response SFT: takes degraded images as input, with the base model’s pre-generated complete responses on the corresponding clean images as supervision targets.

Table 1. Comparison with direct SFT baseline on Qwen3.5-4B. Numbers in parentheses denote changes relative to the base model. CURE consistently outperforms both direct SFT baselines across real-world degradation, synthetic degradation, and clean-image evaluations.

MethodR-Bench (Real-World Degradation) ↑RVE (Synthetic Degradation) ↑GVE (Clean-Image Benchmarks) ↑Avg. ↑
Base (Qwen3.5-4B)63.4973.8977.9171.76
GT-answer SFT46.34 (-17.15)73.21 (-0.68)79.71 (+1.80)66.42 (-5.34)
Clean-response SFT65.89 (+2.40)74.27 (+0.38)78.85 (+0.94)73.00 (+1.24)
CURE67.38 (+3.89)75.25 (+1.36)81.26 (+3.35)74.63 (+2.87)

(1) Contribution of the clean-guided objective (+6.58 on average from GT-answer SFT to Clean-response SFT).
Fine-tuning on very short GQA answers (GT-answer SFT) causes format collapse and impairs instruction following. As a result, the model produces overly brief responses across tasks, dropping R-Bench by 17.15. Replacing ground-truth labels with the model's own clean-image responses (Clean-response SFT) resolves this issue. It yields +6.58 on average over GT-answer SFT and achieves consistent gains over the base model across all three evaluations.

(2) Contribution of the prediction alignment mechanism (+1.63 on average from Clean-response SFT to CURE).
Clean-response SFT and CURE share the exact same degraded inputs, clean-view information source, and training budget. They differ only in supervision target: offline hard labels versus token-level distribution alignment along the model's own trajectory. Under this controlled setup, CURE outperforms Clean-response SFT across all three evaluations (R-Bench +1.49, RVE +0.98, and GVE +2.41). This isolates the independent contribution of the prediction alignment mechanism beyond additional adaptation to degraded images.

AuthorsOpenReview ↗

Author Response – Part 2

Q3: Definition of the term "understanding restoration"

A3: We appreciate the opportunity to clarify this terminology. We use the term "understanding restoration" because image degradation causes a drop in the model's visual understanding. Rather than reconstructing image pixels or only fitting final answers, CURE uses the model's clean-image decision process as a reference to restore its performance on degraded images. Specifically, this is reflected in two aspects:

(1) Restoring model understanding rather than image pixels.
Traditional image restoration optimizes pixel-level or perceptual metrics such as PSNR and SSIM to reconstruct clean images. CURE neither modifies the input image nor requires a restoration module. Instead, it directly restores the model's understanding and decision-making process under degraded inputs.

(2) Using the model's own clean-view decision process as the restoration reference.
Standard fine-tuning on degraded images only optimizes the final answer and does not reference how the model arrives at its response on clean images. In contrast, CURE uses the token-level prediction distributions of the same frozen model on clean images as its restoration reference, bringing the step-by-step decision process under degraded inputs closer to that on clean images.


Q4: Efficiency analysis and scalability

A4: Thank you for this valuable suggestion. We have added analyses and experiments on inference cost, training cost, and scalability. For inference, CURE is identical to the original LVLM at inference, introducing no extra modules or inference memory overhead. For training, CURE requires only a single stage, substantially reducing training cost compared with Robust-R1. Meanwhile, experiments on 4B and 8B models validate the effectiveness of LoRA.

(1) inference cost

We compare the inference efficiency of CURE against the base model, BaryIR + Base, and Robust-R1 on 128 R-Bench samples using a single 64G Ascend 910C NPU.

Table 2. Inference cost comparison on Qwen3.5-4B. CURE introduces no extra module or TTFT overhead, and its higher end-to-end latency is solely due to longer generated responses.

MethodExtra ModuleAvg. Output TokensTTFT (s) ↓End-to-End Latency (s) ↓Relative to Base
Base (Qwen3.5-4B)None59.50.1872.661.00×
CURENone96.00.1954.291.61×
BaryIR + BaseRestoration network56.70.6313.081.16×
Robust-R1None278.30.19212.114.56×
  • Single-model inference (CURE and Robust-R1): Sharing the base architecture, CURE uses the exact same inference memory as the base model, whereas Robust-R1 generates long reasoning chains (278.3 tokens) and incurs 4.56× latency.

  • Restore-then-understand inference (BaryIR + Base): Running BaryIR before the LVLM adds only 0.45 s on low-resolution R-Bench images, but restoring a single 4K image on HRBench-4K takes 30.4 s on average. Because CURE requires no restoration module at inference, it completely avoids this overhead.

(2) Training cost

We compare the training cost of direct SFT, CURE, and the two-stage RL method Robust-R1 on the same Qwen3.5-4B backbone using 8 × 64G Ascend 910C NPUs.

Table 3. Training cost comparison on Qwen3.5-4B. CURE training costs more than direct SFT but is substantially cheaper than SFT+GRPO (Robust-R1).

MethodTraining StageSamplesWall-Clock Time ↓NPU-h / 1K Samples ↓
GT-answer SFTSFT6,4000.72 h0.91h
CUREDistillation6,4001.93 h2.42h
Robust-R1 (same-backbone repro.)SFT + GRPO9,92010.89 h8.78h

Relative to GT-answer SFT, CURE's overhead comes only from single-rollout generation and the clean-branch forward pass. Compared with two-stage Robust-R1, CURE avoids multi-candidate GRPO sampling and cuts wall-clock training time by 5.6× (3.6× lower per 1K samples).

(3) Scalability

We further evaluate CURE with LoRA (r=256) on Qwen3.5-4B and scale it to the larger Qwen3-VL-8B backbone.

Table 4. Scalability to LoRA and a larger backbone. CURE with LoRA consistently improves both 4B and 8B models across all three evaluations.

MethodR-Bench (Real-World Degradation) ↑RVE (Synthetic Degradation) ↑GVE (Clean-Image Benchmarks) ↑
Qwen3.5-4B63.4973.8977.91
CURE (Qwen3.5-4B, full-parameter)67.38 (+3.89)75.25 (+1.36)81.26 (+3.35)
CURE (Qwen3.5-4B, LoRA)66.99 (+3.50)73.91 (+0.02)81.26 (+3.35)
Qwen3-VL-8B62.5372.3278.52
CURE (Qwen3-VL-8B, LoRA)63.54 (+1.01)72.93 (+0.61)79.15 (+0.63)

As shown in Table 4, CURE brings consistent gains across all three benchmarks on Qwen3.5-4B (LoRA) and Qwen3-VL-8B. This confirms that CURE adapts effectively to parameter-efficient settings and larger LVLMs.

Reviewer_Y1nxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CURE, a method to improve LVLM robustness to image degradation via on-policy self-distillation. During training, a trainable branch generates responses conditioned on degraded images, while a frozen copy of the base model provides reference next-token distributions conditioned on paired clean images and the same student-generated prefixes. Training minimizes token-level Jensen–Shannon divergence (JSD) and updates both vision encoder and language decoder. Using 6,400 GQA-derived clean–degraded pairs filtered by mean prediction shift, CURE improves Qwen3.5-4B on R-Bench (real-world degradation), cross-benchmark synthetic degradation, and clean-image understanding in aggregate. Module-freezing and weight-recombination experiments suggest that language-side adaptation accounts for most of the improvement on R-Bench.

Critical strengths
  1. Well-motivated problem framing and a conceptually simple solution. Image restoration often optimizes pixel fidelity or perceptual quality, which need not improve LVLM understanding. CURE uses next-token distributions conditioned on paired clean images to supervise each step of response generation from degraded images. This directly targets downstream understanding and requires only the adapted LVLM at inference.

  2. Comprehensive experimental coverage. CURE improves all nine R-Bench task-severity scores and is evaluated on additional synthetic-degradation and clean-image benchmarks. The freezing and weight-recombination experiments provide useful evidence that language-side adaptation contributes substantially to the observed gains.

Critical weaknesses
  1. Generality across LVLM backbones is not established. All CURE training and ablation experiments use Qwen3.5-4B exclusively. Although the evaluation includes other general-purpose LVLMs for reference, no CURE adaptation is performed on alternative architectures or model scales. This limits confidence in whether the observed gains generalize beyond the specific backbone.

  2. Baseline comparisons lack controlled same-backbone conditions. The robustness methods compared in Tables 1–3 use different native backbones, whereas CURE is adapted from Qwen3.5-4B. The paper does not clarify whether these results are obtained from the original checkpoints or from reimplementations on the same backbone, nor does it provide the corresponding training and evaluation protocols. Consequently, the comparison may conflate differences in backbone capability with the contribution of the proposed training strategy. Please report the backbone, checkpoint, training setup, and evaluation procedure for each baseline and provide matched same-backbone comparisons where feasible.

  3. Novelty relative to prior on-policy self-distillation work is underspecified. The paper cites Vision-OPD (Yuan et al., 2026), which also supervises next-token distributions along student-generated trajectories using privileged visual inputs. CURE applies the same paradigm to clean/degraded pairs with a frozen teacher. The paper should explicitly clarify what is novel beyond this application scenario.

  4. The data-selection procedure is underspecified. The authors select 6,400 pairs from 32,000 candidates, but do not report the thresholds used to exclude examples with excessively high or low prediction shifts. Different filtering criteria may change the distributions of sample difficulty and degradation severity in the training data, limiting reproducibility.

Questions for authors
  1. Does CURE transfer consistently to other LVLM backbones or model scales?
  2. Are the robustness baselines in Tables 1–3 evaluated using their original checkpoints or reimplemented on the same backbone as CURE, and how do matched same-backbone comparisons affect the conclusions?
5 comments
AuthorsOpenReview ↗

Author Response – Part 1

We thank the reviewers for their engagement and thoughtfulness during the review process.

Q1: Generality across different LVLM backbones and model scales

A1: Thank you for this suggestion. This is a great point that prompted us to conduct more comprehensive experiments to support our paper and strengthen our conclusions. Specifically, we added CURE training experiments on Qwen2.5-VL-3B, Qwen3-VL-4B, and Qwen3-VL-8B. The results show that CURE can be reliably transferred to other LVLM backbones and model scales (3B, 4B, and 8B), achieving consistent improvements across all three evaluations: real-world degradations (R-Bench), synthetic degradations (RVE), and clean images (GVE).

Table 1. Generalization of CURE across different LVLM backbones. CURE consistently improves performance across diverse vision-fusion paradigms, LLM attention architectures, and model scales.

Method (Training Setting)Backbone (Vision + LLM)R-Bench (Real-World Degradation) ↑RVE (Synthetic Degradation) ↑GVE (Clean-Image Benchmarks) ↑
Qwen2.5-VL-3BWindow-ViT + Qwen2.5-3B (Full Attn)49.3965.3370.39
CURE (Qwen2.5-VL-3B)51.39 (+2.00)65.52 (+0.19)71.18 (+0.79)
Qwen3-VL-4BSigLIP-2 + Qwen3-4B (Full Attn)60.3170.8277.72
CURE (Qwen3-VL-4B)62.56 (+2.25)71.47 (+0.65)79.23 (+1.51)
Qwen3.5-4BEarly-Fusion ViT + Qwen3.5-4B (Hybrid Linear Attn)63.4973.8977.91
CURE (Qwen3.5-4B)67.38 (+3.89)75.25 (+1.36)81.26 (+3.35)
Qwen3-VL-8BSigLIP-2 + Qwen3-8B (Full Attn)62.5372.3278.52
CURE (Qwen3-VL-8B, LoRA)63.54 (+1.01)72.93 (+0.61)79.15 (+0.63)

As shown in Table 1, these models cover distinct vision-fusion paradigms (window attention with MLP projection, multi-level DeepStack fusion, and native early fusion) and different LLM attention mechanisms (standard full attention vs. 3:1 hybrid linear attention). Across both full-parameter tuning (Qwen2.5-VL-3B, Qwen3-VL-4B, and Qwen3.5-4B) and LoRA tuning (Qwen3-VL-8B), CURE outperforms every base model on all benchmarks. This confirms that CURE generalizes across alternative architectures and model scales.

Q2: Checkpoints and training setups of the baselines, and same-backbone comparisons

A2: We apologize for not clarifying this in the paper. The robustness baselines in the paper were evaluated using their official released checkpoints. We will also explicitly add their experimental setups in the revision. In addition, we conducted controlled same-backbone comparisons to more accurately evaluate the effectiveness of our proposed method. The results confirm that CURE consistently outperforms the baseline across all three evaluations under the same backbone.

(1) Backbones and setups of the robustness baselines. In Tables 1–3, RestoreVAR and BaryIR serve as image-restoration front ends. Their restored images are fed directly into Qwen3.5-4B (or CURE). The native backbones and architectures of the other four robustness baselines are listed in Table 2.

Table 2. Native backbones and architectures of the robustness baselines.

BaselineNative Backbone & Architecture
Robust CLIPLLaVA-v1.5-7B (with OpenAI CLIP ViT-L 224px as vision encoder)
Robust LLaVALLaVA-v1.5-7B (with AdvXL-Giant and AdvXL-Huge as vision encoders)
Robust-R1Qwen2.5-VL-3B
Robust-U1BAGEL-7B-MoT (unified visual understanding and generation model)

Among them, Robust CLIP and Robust LLaVA rely on specialized adversarially trained vision encoders, while Robust-U1 relies on a dedicated unified understanding-and-generation architecture.

(2) Matched same-backbone comparisons. To remove the influence of backbone differences, we directly compare Robust-R1 and CURE on two identical backbones.

Table 3. Matched same-backbone comparisons between Robust-R1 and CURE. "*" indicates results evaluated from official released checkpoints. CURE consistently outperforms Robust-R1 on all three benchmarks under both backbones.

MethodR-Bench (Real-World Degradation) ↑RVE (Synthetic Degradation) ↑GVE (Clean-Image Benchmarks) ↑
Qwen2.5-VL-3B*49.3965.3370.39
Robust-R1 (Qwen2.5-VL-3B)*50.42 (+1.03)64.93 (-0.40)70.48 (+0.09)
CURE (Qwen2.5-VL-3B)51.39 (+2.00)65.52 (+0.19)71.18 (+0.79)
Qwen3.5-4B*63.4973.8977.91
Robust-R1 (Qwen3.5-4B)62.06 (-1.43)74.04 (+0.15)78.49 (+0.58)
CURE (Qwen3.5-4B)67.38 (+3.89)75.25 (+1.36)81.26 (+3.35)

As shown in Table 3, CURE outperforms Robust-R1 across all three evaluations on both backbones. This confirms that the performance gains of CURE come from our method itself rather than differences in backbone capability.

AuthorsOpenReview ↗

Author Response – Part 2

Q3: Novelty of CURE

A3: Thank you for this comment. CURE starts from the problem of image degradation, and on-policy distillation is the technique we adopt to address it. Both CURE and Vision-OPD follow the training objective of on-policy distillation.

(1) Existing approaches to degradation are not aligned with model understanding

LVLMs are increasingly becoming important downstream consumers of degraded images. The two main lines of existing methods do not directly optimize LVLM understanding. Image restoration optimizes pixel or perceptual quality, which does not guarantee better understanding and even lowers the base model's score on R-Bench. Meanwhile, robust fine-tuning aligns visual features or final answers, without constraining the intermediate token-by-token decisions altered by degradation.

(2) Shifting the restoration target from pixels to model decisions

Based on this analysis, CURE restores the model's decisions rather than image pixels. Using a frozen clean-image branch as the teacher, CURE aligns the student's next-token distributions along trajectories generated on degraded images. This requires no human annotations or external teacher, and needs no restoration module at inference.

The main difference from Vision-OPD lies in the source of the teacher's information. In Vision-OPD, the teacher sees a region cropped from the student's image. The relevant information is already contained in the student's input, and the goal is to improve the model's ability to focus. In contrast, degradation can weaken or remove part of the visual evidence. The student may not have access to the information that the teacher relies on. In this more challenging setting, CURE still improves R-Bench from 63.49 to 67.38 and also improves performance on clean images.

(3) Relationship between understanding restoration and image restoration

CURE can be used alone or as the understanding stage of a restoration pipeline. On real-world degradations (R-Bench), both restoration front ends lower the score of the base model, and CURE alone performs best. On synthetic degradations (RVE), BaryIR improves the base model, and combining it with CURE brings further gains. With either front end, replacing the base model with CURE also improves the results.

(4) Understanding restoration mainly happens on the language side

Prior robustness methods mostly strengthen the vision encoder. Our module-freezing and weight-recombination experiments on Qwen3.5-4B show a different picture. On R-Bench, training only the language model retains about 89% of the full gain, while training only the vision encoder gives limited improvement. Combining the trained language model with the original vision encoder still retains most of the gain.


Q4: Thresholds for data selection and their effect

A4: Thank you for this helpful suggestion. In our pipeline, prediction-shift filtering serves merely as a lightweight sanity check to remove extreme outliers (e.g., δt≈0\delta_t \approx 0 with negligible degradation effect, or rare tail cases), rather than relying on a carefully tuned threshold. After this basic filtering, we simply draw 1,280 random pairs per degradation operator (N=6,400N = 6,400).

To verify whether this step affects performance or reproducibility, we retrained CURE using unfiltered, operator-balanced random samples across 3 random seeds (6,400 pairs), and at a larger scale (12,800 pairs).

Table 4. Comparison between filtered and unfiltered random training sets (Qwen3.5-4B). Numbers in parentheses denote changes relative to the base model.

MethodR-Bench (Real-World Degradation) ↑RVE (Synthetic Degradation) ↑GVE (Clean-Image Benchmarks) ↑
Base (Qwen3.5-4B)63.4973.8977.91
CURE (Filtered, 6,400 pairs, reported)67.38 (+3.89)75.25 (+1.36)81.26 (+3.35)
CURE (Unfiltered, 6,400 pairs, 3 seeds)67.30 (+3.81) ± 0.9874.91 (+1.02) ± 0.5381.36 (+3.45) ± 0.84
CURE (Unfiltered, 12,800 pairs)67.91 (+4.42)74.51 (+0.62)80.38 (+2.47)

As shown in Table 4, unfiltered random sampling across 3 seeds closely matches the reported results and scaling to 12,800 unfiltered pairs further improves R-Bench to 67.91. Thus, CURE does not rely on specific filtering thresholds and is easily reproducible. We will clarify this in the revision and release our data-construction scripts and sample indices.

Thanks for the additional experiments and for clarifying the baseline setup. This addresses most of my concerns. Please include the results across Qwen model series and sizes, along with the same-backbone comparisons with Robust-R1, in the final paper.

I still have a question about Robust-U1. It scores 53.45 in Table 1, compared with 73.98 in the original paper. What might explain this 20.53-point gap? Did you follow the original procedure of first restoring the image and then giving the model both the corrupted and restored images? I also noticed that Robust-U1 is missing from Table 2. Could you explain why and report results in both tables using its original inference procedure, with the same test data and scoring rules ?

I’m open to raising my score based on your response.

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

in this paper, authors aim to make the VLM be robust to degraded images with applying any preprocessing image restoration. Therefore, they propose CURE. it is sort of self-distillation method. A teacher VLM answer question with clean image. the student VLM learn those teacher's labels on degraded image. during training, no gt answers are used. Several typical degradation types discussed like gaussian blur, compression degradation, and contrast, etc. The experiment shows that the trained CURE can brings improvement compared to the base models but gain little improvement compare to the strongest and most naive baselines.

Critical strengths

(+) the problem settings is valuable as most of real-world images have flaw and degradation, which challenges the VLM and QA.

(+) the solution is straightforward and easy to implement, which indeed brings incremental improvement.

Critical weaknesses

(-) it seems a neccesary comparison is missing, which is degraded image as input and GT answer as direct SFT. Author simply discuss that, but no comparison is provided explicitly.

(-) the diversity and realistic of the synthetic degradation raise my concern about syn2real. In current settings, all those degradations are using pretty naive degradation models like contrast, Gaussian blur, Gaussian noise, JPEG, etc. those usually cannot reflect the degradation in real world, which limits the value of the proposed method. For example, can such model trained on gaussian blur understand the spatially varying blur like defocus blur and severe atmospheric turbulence? It will be interesting to see that VLM model show generalization capability on diverse degradation types. Train a easy blur but work on complex blur. There could be interesting insights.

(-) the comparison and improvement is over-claim. the most important baseline is restoration-based BaryIR because it is a naive solution against CURE. in the table, CURE+BaryIR should be removed or at least reader should not focus on. If we only focus on restoration-based BaryIR vs. CURE. the improvement is not impressive at all. CURE even falls behind on some benchmarks, which raises my concern about the effectiveness and necessity of the proposed method. Can author justify that?

Questions for authors

I have listed all of my concerns in the weakness part.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for the valuable feedback. We address each question point by point below.

Q1: Comparison with direct SFT on degraded images using ground-truth answers

A1: Thank you for this suggestion. This is a valuable comment that prompted us to conduct more comprehensive experiments to support our paper and strengthen our conclusions. We have added two SFT baselines on Qwen3.5-4B, and CURE outperforms both SFT baselines across all three evaluation settings: real-world degradations, synthetic degradations, and clean images.

Both SFT baselines use the exact same training data size, number of training steps, and learning rate as CURE:

  • GT-answer SFT: takes degraded images as input, with GQA ground-truth answers as supervision targets;
  • Clean-response SFT: takes degraded images as input, with the base model’s pre-generated complete responses on the corresponding clean images as supervision targets.

Table 1. Comparison with direct SFT baselines on degraded images (Qwen3.5-4B). Numbers in parentheses denote changes relative to the base model. CURE consistently outperforms both direct SFT baselines across real-world degradation, synthetic degradation, and clean-image evaluations.

MethodR-Bench (Real-World Degradation) ↑RVE (Synthetic Degradation) ↑GVE (Clean-Image Benchmarks) ↑
Base (Qwen3.5-4B)63.4973.8977.91
GT-answer SFT46.34 (-17.15)73.21 (-0.68)79.71 (+1.80)
Clean-response SFT65.89 (+2.40)74.27 (+0.38)78.85 (+0.94)
CURE67.38 (+3.89)75.25 (+1.36)81.26 (+3.35)

As Table 1 shows, GT-answer SFT slightly improves GVE (+1.80) but severely degrades R-Bench (-17.15). Clean-response SFT serves as a stronger baseline by improving all three evaluations, yet CURE still achieves consistently larger gains on every benchmark.

Notably, Clean-response SFT and CURE both learn from the model's own behavior on clean images, and they differ only in how this supervision is applied. Clean-response SFT fits offline pre-generated responses as hard labels. In contrast, CURE aligns the next-token prediction distributions along prefixes generated by the student on degraded images. The clear gap between them confirms that CURE's gains do not merely come from additional training on degraded images.

AuthorsOpenReview ↗

Rebuttal by Authors

Q2: Generalization from simple synthetic to complex real-world degradations

A2: Thank you for guiding us to expand our experiments and analysis on generalization to complex degradations. Following your suggestion, we ran additional training experiments using single, simple degradation types. The results show that synthetic training with CURE on simple degradations (e.g., Gaussian blur) can transfer to more complex real-world degradations (e.g., Motion blur and Lens blur).

For each of the five degradation operators (Contrast, Gaussian blur, Gaussian noise, JPEG, and Pixelate), we trained a CURE model that uses only a single operator. All other settings are the same as in mixed training. We examine the results from two angles.

(1) Syn2Real generalization: CURE trained on synthetic degradations generalizes well to degradations in the real world (R-Bench).

Training uses simple synthetic operators, but our evaluation also includes R-Bench, a benchmark of real captured images with degradations from the entire imaging pipeline. Table 2 shows that CURE improves on both the real-world setting (R-Bench) and the synthetic setting (RVE). So the gain does not depend on one particular operator.

Table 2. Results of single-degradation training on the three evaluations (Qwen3.5-4B). Training on a single simple operator achieves performance close to mixed training across real-world, synthetic, and clean-image evaluations.

Method (Training Setting)R-Bench (Real-World Degradation) ↑RVE (Synthetic Degradation) ↑GVE (Clean-Image Benchmarks)↑
Base (Qwen3.5-4B)63.4973.8977.91
CURE (trained only on Contrast)67.1274.4780.45
CURE (trained only on Gaussian blur)66.9775.0682.41
CURE (trained only on Gaussian noise)67.2174.7080.45
CURE (trained only on JPEG)66.5174.6479.99
CURE (trained only on Pixelate)66.8474.7480.61
CURE (trained on all 5 operators)67.3875.2581.26

(2) Train on an easy blur, work on complex blurs: CURE trained only on Gaussian blur understands spatially varying blurs.

R-Bench contains complex blur types that are not seen during training. To look at this more closely, we report the per-category results on R-Bench in Table 3.

Table 3. Per-category R-Bench results of single-degradation training (Qwen3.5-4B). Numbers in parentheses are changes relative to the base model. CURE trained on simple synthetic degradations transfers to unseen, complex real-world blurs.

Method (Training Setting)Motion blur (Real-World Blur) ↑Lens blur (Real-World Blur) ↑Lens obstacle (Real-World Occlusion) ↑
Base (Qwen3.5-4B)69.3357.1964.00
CURE (trained only on Contrast)73.93 (+4.60)61.41 (+4.22)63.41 (-0.59)
CURE (trained only on Gaussian blur)73.48 (+4.15)65.26 (+8.07)62.81 (-1.19)
CURE (trained only on Gaussian noise)72.81 (+3.48)61.26 (+4.07)59.56 (-4.44)
CURE (trained only on JPEG)73.19 (+3.86)64.37 (+7.18)65.63 (+1.63)
CURE (trained only on Pixelate)75.78 (+6.45)58.67 (+1.48)66.52 (+2.52)
CURE (trained on all 5 operators)72.44 (+3.11)64.89 (+7.70)65.56 (+1.56)

Motion blur and Lens blur are more complex, spatially varying blurs. Even so, CURE (trained only on Gaussian blur) improves on these two categories by 4.15 and 8.07 points, respectively. Based on these results, we believe that the synthetic training of CURE transfers, to a certain extent, to real degradations, and that training on simple blur can transfer to more complex blur.

AuthorsOpenReview ↗

Rebuttal by Authors

Q3: Direct comparison with BaryIR

A3: Thank you for this valuable comment. We agree that the restoration-based BaryIR + Base is the most direct baseline and should be compared directly with CURE. In this direct comparison, their performance difference stems from the degradation characteristics of different datasets: BaryIR only excels on simple synthetic degradations (RVE), whereas CURE significantly outperforms BaryIR on more critical real-world degradations (R-Bench) and clean images (GVE), achieving consistent gains across all three evaluations. Table 4 lists Base, BaryIR + Base, and CURE across the three evaluations, along with BaryIR + CURE as a supplementary observation.

Table 4. Comparison between BaryIR and CURE on Qwen3.5-4B. Numbers in parentheses are changes relative to the base model. CURE outperforms BaryIR on real-world and clean benchmarks while improving all three evaluations.

MethodR-Bench (Real-World Degradation) ↑RVE (Synthetic Degradation) ↑GVE (Clean-Image Benchmarks) ↑
Base (Qwen3.5-4B)63.4973.8977.91
BaryIR + Base (Restoration + LVLM)62.72 (-0.77)76.20 (+2.31)76.07 (-1.84)
CURE (Single LVLM, Ours)67.38 (+3.89)75.25 (+1.36)81.26 (+3.35)
BaryIR + CURE66.47 (+2.98)76.86 (+2.97)80.49 (+2.58)

(1) The performance difference stems from degradation complexity.

As an end-to-end restoration network, BaryIR fits the simple synthetic degradations in RVE well (76.20), but degrades performance on complex real-world degradations (R-Bench, -0.77) and clean images (GVE, -1.84). In contrast, CURE improves on them by +3.89 and +3.35, respectively.

(2) Effectiveness and necessity of CURE.

  • Effectiveness: CURE consistently improves all three evaluations when used alone, and further improves over BaryIR + Base across all benchmarks when combined (BaryIR + CURE).
  • Necessity: Pixel restoration alone hurts performance on real-world and clean images and requires running an extra restoration network before every inference; CURE restores understanding with a single model and no extra inference cost.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global Response by Authors

Dear Area Chairs and Reviewers,

We sincerely thank all reviewers for their time and constructive feedback. We are encouraged that the reviewers highlighted several key strengths of our work, including the practical importance of improving LVLM understanding under real-world image degradation (Reviewers fkSd, Y1nx, and qcRn), the simplicity of CURE’s clean-guided, token-level supervision and its module-free inference (Reviewers fkSd, Y1nx, and qcRn), and the value of our evaluation and analysis showing that language-side adaptation accounts for most of the gains (Reviewers Y1nx and qcRn).

In response to the reviewers’ questions, our additional experiments and clarifications show that:

  • CURE outperforms direct SFT on degraded images, isolating the gains of clean-guided supervision and token-level prediction alignment (Reviewers fkSd and qcRn);
  • CURE generalizes consistently across alternative LVLM backbones, larger model scales, and parameter-efficient LoRA adaptation (Reviewers Y1nx and qcRn);
  • Matched same-backbone comparisons confirm that CURE outperforms Robust-R1, with all official baseline checkpoints clarified (Reviewer Y1nx);
  • Training on simple synthetic degradations transfers to complex real-world blurs, and CURE outperforms restoration-based BaryIR on real-world and clean benchmarks (Reviewer fkSd);
  • CURE differs conceptually from Vision-OPD and standard fine-tuning by performing "understanding restoration" on corrupted visual evidence (Reviewers Y1nx and qcRn);
  • Unfiltered random training pairs across multiple seeds reproduce our results without relying on data-selection thresholds (Reviewer Y1nx);
  • CURE adds no extra modules, inference memory, or TTFT overhead, while requiring much lower training cost than Robust-R1 (Reviewer qcRn).

We hope these additional results and clarifications address the reviewers’ concerns. We appreciate their feedback and look forward to further discussion.

Best regards,
Authors

Dear Reviewers,

As the discussion period is coming to a close, I would like to encourage everyone to review the authors' responses and engage in any remaining discussions. Please indicate whether your concerns have been adequately addressed and update your scores if your assessment has changed.

Your active participation is greatly appreciated and will be valuable in reaching a fair and well-informed decision.

Thank you for your time and effort!

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.