- 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
-
The paper identifies that pixel-level restoration quality does not necessarily correlate with downstream LVLM understanding, which is an important issue for multimodal robustness.
-
The clean-guided distribution alignment objective is intuitive and technically reasonable.
-
The paper evaluates multiple degradation settings and provides useful analysis of visual and language adaptation contributions.
Critical weaknesses
-
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.
-
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.
-
The term "understanding restoration" is ambiguous and requires further justification.
-
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
-
How much of the performance improvement comes from the proposed clean-guided objective itself, compared with simply fine-tuning the LVLM on degraded images?
-
Why is the proposed approach considered "understanding restoration" rather than general degraded-image robustness fine-tuning?
-
What is the computational overhead of CURE compared with existing restore-then-understand pipelines?
-
How scalable is CURE to larger LVLMs or parameter-efficient adaptation settings?
2 comments
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.
| Method | Extra Module | Avg. Output Tokens | TTFT (s) ↓ | End-to-End Latency (s) ↓ | Relative to Base |
|---|---|---|---|---|---|
| Base (Qwen3.5-4B) | None | 59.5 | 0.187 | 2.66 | 1.00× |
| CURE | None | 96.0 | 0.195 | 4.29 | 1.61× |
| BaryIR + Base | Restoration network | 56.7 | 0.631 | 3.08 | 1.16× |
| Robust-R1 | None | 278.3 | 0.192 | 12.11 | 4.56× |
-
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).
| Method | Training Stage | Samples | Wall-Clock Time ↓ | NPU-h / 1K Samples ↓ |
|---|---|---|---|---|
| GT-answer SFT | SFT | 6,400 | 0.72 h | 0.91h |
| CURE | Distillation | 6,400 | 1.93 h | 2.42h |
| Robust-R1 (same-backbone repro.) | SFT + GRPO | 9,920 | 10.89 h | 8.78h |
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.
| Method | R-Bench (Real-World Degradation) ↑ | RVE (Synthetic Degradation) ↑ | GVE (Clean-Image Benchmarks) ↑ |
|---|---|---|---|
| Qwen3.5-4B | 63.49 | 73.89 | 77.91 |
| 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-8B | 62.53 | 72.32 | 78.52 |
| 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.
Author Response – Part 1
Thank you for your valuable and constructive feedback.
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.
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:
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.
(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.