acceptodds
Under review as a conference paper at ICLR 2027

LVP-DiffIR: Distilling Large-Model Trajectories for Thermal Infrared Image Restoration

Abstract

Infrared image enhancement faces a persistent trade-off: compact networks trained on synthetic pairs are efficient but overfit the simulator, while large generative models are accurate yet computationally expensive at inference.We propose Large-model Visual Prior Diffusion Image Restoration (LVP-DiffIR), a large-to-small framework that distills an offline large-model iterative enhancement trajectory into a compact five-step student.The student copies both the first restore, which removes the dominant degradations, and the later residual refinement of edges and contrast.To save large-model inference time during distillation, we adapt the student with Median-Ratio Adaptation (MRA) on unlabeled real infrared: an IQA memory bank and a per-source median-ratio backward gate, built on an EMA teacher–student pair. The adapted student further outperforms pure trajectory distillation.We further introduce InfraRS, to our knowledge the highest-resolution public remote sensing dataset for infrared image enhancement.Distilled from offline RAR trajectories and then adapted with MRA on unlabeled real infrared, LVP-DiffIR reaches state-of-the-art no-reference perceptual scores on held-out real images, including InfraRS, without paired real-world ground truth on those captures.Our code will be released soon.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Enhancement

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_oeg4OpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LVP-DiffIR, a large-to-small infrared image restoration framework that transfers the restoration capability of a large iterative restoration model into a compact student network. The proposed method consists of two stages. First, the restoration trajectory generated by an offline RAR model is distilled into a five-step compact diffusion restoration model. Second, the distilled model is further adapted using unlabeled real infrared images through Median-Ratio Adaptation (MRA). In addition, the authors introduce InfraRS, a high-resolution infrared remote sensing dataset for evaluating infrared image enhancement. The experiments demonstrate improved no-reference restoration quality on several infrared benchmarks.

Critical strengths
  1. The paper identifies an important limitation of existing restoration models: compact networks trained on synthetic degradation pairs often suffer from domain gaps, while large iterative restoration models are computationally expensive. The proposed method provides a potential solution.
  2. The paper introduces InfraRS, providing a new dataset for evaluating infrared enhancement under realistic conditions.
Critical weaknesses
  1. The proposed framework relies on restoration trajectories generated by RAR. However, the paper does not investigate how sensitive the student model is to the teacher quality.
  2. Although the proposed method reduces inference cost by replacing the original iterative large restoration model with a compact student network, the overall training pipeline still requires generating restoration trajectories from the large teacher model. A detailed comparison between training and inference cost would better demonstrate the practical value of the proposed framework.
  3. The paper does not sufficiently analyze whether the improvement mainly comes from the proposed adaptation mechanism or simply from additional fine-tuning on more real infrared samples.
Questions for authors
  1. Does a weaker teacher still provide useful trajectories?
  2. How much performance improvement comes from RAR itself rather than the proposed distillation strategy?
  3. How does the proposed trajectory distillation compare with directly distilling the final restoration output of the teacher model?
  4. What is the computational cost of generating the offline restoration trajectories compared with training a conventional restoration model?
Reviewer_BmLFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes LVP-DiffIR for infrared image enhancement, which transfers the restoration ability of a large iterative restorer (RAR) into a compact five-step diffusion student. In the first stage, RAR is run offline on a small set of images, and the student is trained on the resulting trajectory In the second stage, the student is adapted on unlabeled real infrared images through an EMA teacher-student scheme, where a per-image memory bank keeps a pseudo-label only if it improves both MUSIQ and NIQE, and a median-ratio gate decides whether a sample produces a gradient. A Dual Canny loss is added to suppress spurious weak edges. The authors also release InfraRS, a 10 m satellite infrared dataset.

Critical strengths
  1. Distilling an expensive test-time Restore-Assess-Repeat loop into a five-step student using only 150 teacher trajectories is a practical direction for infrared applications. The two-center schedule is motivated by an analysis of how the teacher trajectory behaves (Fig. 3) and is supported by its ablation (Table 6).

  2. InfraRS covers satellite infrared imagery, which differs clearly from the ground-level data that existing infrared restorers are trained on, and could be useful to the community if properly documented.

Critical weaknesses
  1. The evaluation is largely circular. Stage 2 selects pseudo-labels and gates gradients with MUSIQ and NIQE, and the main results are reported on the same kind of NR-IQA metrics, while PSNR/SSIM are omitted on HM-TIR even though paired GT is available. Fidelity results, and clarification of whether test images were used for checkpoint selection (A.1), are needed to show that the student restores rather than optimizes the metrics.

  2. The comparison is not consistent. QualiTeacher is excluded for edge artifacts, while Figs. 9-11 admit that frames with similar ink-like outlines were left out of the paper's own results. The baselines also never see InfraRS or M3FD during training, so part of the generalization gain comes from domain exposure.

  3. Efficiency is the main motivation, but it is never measured. There is no report of inference time, FLOPs or parameters against RAR and the compact baselines, nor of the cost of generating the teacher trajectories, and the radar chart in Fig. 1 has no numerical basis.

  4. Closely related baselines are missing. Few-step and one-step diffusion restorers obtained by distillation (e.g., SinSR, OSEDiff) are the most direct competitors for a five-step student, and general diffusion restorers such as DiffBIR are also absent, so it is hard to tell how much the RAR trajectory itself contributes.

Questions for authors

Beyond the requests above: the full-model numbers in Table 4 differ from those in Tables 2 and 3, and results come from a single seed while several ablation gaps are around one MUSIQ point; please explain the mismatch and report variance. For InfraRS, please state the spectral band, whether 10 m is the native resolution, and the release license, since this affects the "highest-resolution" claim.

Reviewer_CXf1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LVP-DiffIR is a two-stage framework for infrared image restoration. It first distills a large model’s iterative restoration process into a compact five-step student. It then adapts the student to unlabeled infrared images using a quality-filtered pseudo-label bank, a median-ratio training gate, and an edge regularizer. The paper also introduces InfraRS, a high-resolution infrared remote sensing dataset. Experiments show improvements over compact baselines on most no-reference quality metrics, although restoration fidelity and computational efficiency need further validation.

Critical strengths
  1. The framework uses only 150 RAR triplets for initial distillation and then adapts to additional unlabeled images without further RAR inference, reducing the need for costly teacher-generated supervision.
  2. InfraRS contains 6,731 infrared remote sensing patches at a ground sampling distance of 10m, providing a useful resource for studying restoration in diverse overhead scenes.
Critical weaknesses
  1. The benefit of trajectory distillation remains unclear. Trajectory distillation appears to be the paper’s main methodological contribution. However, Table 6 only compares different uses of intermediate teacher outputs, while the “w/o trajectory distillation” variant in Table 4 is insufficiently specified. Please clarify whether this variant uses only the input x and final teacher output y_last. A controlled comparison that removes y_1 while keeping the backbone, prediction objective, and training budget unchanged would help isolate the contribution of intermediate trajectory information.
  2. Missing full-reference evaluation on HM-TIR. HM-TIR provides paired test references, but the paper omits PSNR and SSIM because paired references are not used during training. This justification is insufficient, as the training supervision does not prevent evaluation against held-out ground truth. Reporting PSNR, SSIM, and LPIPS alongside no-reference metrics would help assess whether the perceptual improvements come at the cost of reconstruction fidelity.
Questions for authors
  1. Why do the full-model results differ between the main comparison tables and the ablation table? For example, the InfraRS MUSIQ score is 54.74 in Table 3 but 55.03 in Table 4, with similar discrepancies on HM-TIR and M3FD. Please clarify whether these results use different checkpoints, evaluation subsets, or random seeds, and ensure that ablation gains are measured under a consistent protocol.
  2. Since computational efficiency is a central motivation, could the authors report inference latency, parameter counts, and peak GPU memory for LVP-DiffIR, RAR, and compact baselines under the same hardware and input resolution? Five sampling steps alone do not establish a practical speed advantage.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.