acceptodds
Under review as a conference paper at ICLR 2027

GranFlow: Granularity-Matched Conditioning for Degradation-Robust Infrared-Visible Fusion via Rectified Flow

Abstract

Compound degradations in real-world infrared–visible imaging undermine the reliable integration of complementary information, making degradation-robust fusion essential for preserving scene structure and restoring visual appearance. These degradations affect image content in fundamentally different ways: noise and stripe artifacts introduce spurious artifacts, blur attenuates high-frequency detail, and haze induces spatially varying contrast loss together with broader shifts in image statistics. Distinguishing frequency-dependent corruption from broader distributional shifts can therefore help balance artifact suppression, complementary detail preservation, and global appearance restoration. We propose GranFlow, a Rectified Flow fusion framework combining a wavelet-domain gated module that adaptively reweights modalities for frequency-selective corruptions, and a scale-aware global conditioning mechanism for distributional shifts. We find that this global signal captures sample-dependent variation in latent displacement scale, producing scale-adaptive modulation without degradation labels. Experiments across single and compound degradations demonstrate that degradation-adaptive modality reweighting and scale-adaptive latent refinement improve structural fidelity and visual appearance while preserving clean-input fusion quality and benefiting downstream object detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Degraded Multi-modal Image Fusion, Rectified Flow, Degradation-Robust Image Restoration, Wavelet-Based Conditioning, Conditional Generative 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_rs2eOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes GranFlow, a two-stage Rectified Flow framework for infrared-visible image fusion under degraded inputs. In Stage I, wavelet-based subband statistics and CNN features are used to condition modality-specific encoding and dynamically gate the fusion of infrared and visible latent features. In Stage II, a conditional DiT refines the fused latent representation through Rectified Flow, using a sample-dependent latent displacement scale as an additional conditioning signal. The model is trained with synthetic noise, blur, stripe, and haze degradations and is evaluated on standard fusion benchmarks, compound degradation settings, additional datasets with naturally occurring or real-world degradations, and downstream detection or segmentation tasks. The main claim is that matching conditioning granularity to degradation characteristics improves degradation-robust fusion.

Critical strengths
  1. The paper evaluates degradation-robust fusion beyond the clean-input setting. It considers single degradations, compound degradations, and degradation-free inputs across multiple datasets. It also evaluates the fused outputs on object detection and semantic segmentation, providing task-level validation in addition to conventional fusion metrics.
  2. The framework separates two different forms of conditioning within a single fusion pipeline. The Stage-I wavelet branch extracts modality-specific subband information and produces spatial-channel fusion gates, while the Stage-II Rectified Flow branch refines the fused latent using a sample-dependent displacement scale. This separation gives the proposed architecture a clear functional organization, and the reported comparisons with RF-Fusion and the module ablations indicate that the complete model can improve performance under the tested settings.
Critical weaknesses
  1. The central motivation is not sufficiently convincing. The paper treats differences in frequency-domain and statistical characteristics across degradations as the main motivation. However, this observation is not by itself a compelling research gap, since prior fusion methods have already exploited degradation-specific frequency information, such as ControlFusion. The Introduction should more carefully analyze where existing degradation-robust fusion methods fail and why their conditioning strategies are inadequate. The broad schematic in Figure 1(c) does not provide enough support for this argument.
  2. The proposed degradation formulation has a limited scope, and its real-world robustness is not demonstrated. The design is mainly motivated by noise, blur, stripe artifacts, and haze, but it is unclear whether the same frequency-statistical formulation applies to common conditions such as low-light and over-exposure. Most controlled experiments use synthetic degradations, while the real-world experiments do not provide degradation-specific analysis. In particular, Figure 9 only reports object detection results on LLVIP and does not show that the method can effectively handle low-light degradation.
  3. The methodological contribution appears largely incremental. The proposed framework combines several established components, including wavelet decomposition, statistical and CNN descriptors, conditional modulation, gated fusion, VAE-based latent modeling, DiT, Rectified Flow, and scale conditioning. Although some components are redesigned for infrared-visible fusion, the paper does not clearly identify a methodological principle that goes beyond the combination of existing techniques.
  4. The paper does not directly verify its central adaptive-fusion mechanism. The main claim is that the model adapts the relative contributions of the infrared and visible modalities according to the degradation type and its characteristics. However, the paper does not visualize or quantitatively analyze the learned modality gates, subband weights, or scale-dependent modulation.
Questions for authors
  1. Which specific failure modes of existing degradation-robust fusion methods are not addressed by prior frequency-aware or degradation-conditioned approaches? Please clarify why the proposed granularity-matched conditioning is necessary.
  2. What evidence supports applying the proposed frequency-statistical formulation to degradations beyond noise, blur, stripe artifacts, and haze, such as low-light and overexposure? Since Figure 9 reports only detection results on LLVIP, what evidence supports the claim that GranFlow can handle low-light degradation?
  3. Which component or learning principle of GranFlow goes beyond the combination of wavelet decomposition, conditional modulation, gated fusion, VAE, DiT, Rectified Flow, and scale conditioning? Please clarify the specific methodological contribution that distinguishes GranFlow from existing frequency-aware, degradation-aware, and Rectified Flow-based fusion methods.
  4. Can the authors provide direct visualization or quantitative analysis of the learned modality gates, subband weights, and scale-dependent modulation under different degradation types?
7 comments
AuthorsOpenReview ↗

Response to Reviewer rs2e (Part.1)

Thank you for your comments. We have clarified the design rationale and methodological contribution, expanded the evaluation under unseen degradations, and added capacity-comparable conditioning varients comparisons and direct analyses of the conditioning mechanisms. The relevant evidence is presented in Sec. 4.5 (p. 9, Line 451), App. 11 (p. 24, Line 1262), and App. 12 (p. 25, Line 1346). Together, these results support the rationale for our design and demonstrate its effectiveness under the evaluated conditions. We address each point below.


W1/Q1 — Design Motivation and Limitations of Existing Fusion Methods

Different degradations differ in their statistical characteristics, including frequency content and latent displacement. These differences reflect the intrinsic relationships in image restoration under different degradations and can be used as the conditioning signals to achieve adequate restoration of complexly degraded images while preserving their useful information. We therefore use them to motivate our design.

Existing methods for degraded infrared-visible image fusion still have design limitations that leave some restoration needs insufficiently addressed. Methods such as AMG-Fuse use modality-contribution masks to guide degradation suppression and cross-modal interaction. However, masks based on intensity-mixing relationships struggle to separate true colors, differences between modalities, and degradation effects. This leads to substantial color bias in the fused results, as shown in the haze example of Fig. 4. All-in-one methods such as TG-ECNet use experts or prompts within a shared architecture to handle different degradation types and strengths. This limits how well the model can meet the restoration needs of an individual sample. Methods such as ControlFusion use frequency conditions and text-based modeling to describe degradation globally. However, a global condition alone cannot provide local information aggregation within the image. These limitations make it important to fully use frequency information for restoration at the degradation-type level and latent displacement scale for refinement at the sample level. GranFlow follows this design: modality-specific frequency features guide restoration and fusion, while the current latent displacement scale guides further refinement for each sample. The other two reviewers also recognize this motivation: Reviewer vAJJ describes our design as "well motivated," and Reviewer K4b3 finds the motivation "reasonable." The analysis above explains how this motivation leads to our two-stage design.

We have also clarified the different pipeline types in Fig. 1: Fig. 1(c) shows guidance from input-derived prompts, while Fig. 1(d) specifies our spatial-channel gating and displacement-conditioned refinement. This distinction makes the motivation behind our design clearer.


W2/Q2 — Unseen Degradations and Transfer to Exposure Changes

Table 19 directly evaluates fusion under low light and overexposure. The fusion results measure image quality under these two exposure changes. We also test stronger noise and blur, a different infrared-noise setting, and local degradation. Tables 18–21 give the individual results, Fig. 11 shows example outputs, and Table 6 gives the averages.

Test settingMSE ↓PSNR ↑MS-SSIM ↑
Extreme noise0.034315.30710.6618
Extreme blur0.035415.19850.6374
Low light0.034415.20920.6913
Overexposure0.051413.42900.6668
Noisy infrared0.034315.38520.6788
Local noise0.034515.35940.6783
Local blur0.033615.48460.6972

GranFlow has the best MSE, PSNR, and MS-SSIM under both low light and overexposure among the compared methods. It also has the best SCD under low light and QNCIE under overexposure. Figure 11 shows clearer targets and background structure in the low-light example. In the overexposure example, the bright area spreads less and nearby content remains visible. These results directly support the method's ability to fuse images under the tested exposure changes. GranFlow uses image features rather than a fixed set of degradation labels. The wavelet statistics and learned features come from the input images. The gate also uses local latent features, while the rectified flow network uses the fused latent and its displacement. Neither stage needs a degradation label at inference. The model can therefore respond to changes in brightness, contrast, and structure without assigning the image to a named degradation class first. Haze, low light, and overexposure have different physical causes. However, handling them can involve some shared needs, such as adjusting contrast and intensity while keeping useful structure. These shared needs motivate transfer to exposure changes, and Table 19 directly demonstrates this transfer in the tested settings.

AuthorsOpenReview ↗

Response to Reviewer rs2e (Part.2)

Extreme noise and blur exceed the training strengths. The infrared-noise test differs from the training stripe setting: training uses independent Gaussian noise at 15/25515/255–20/25520/255 together with noise shared within columns, while the new test uses grayscale Gaussian noise at 25/25525/255. This tests a change in both infrared-noise structure and strength. The local tests restrict noise or blur to part of the image. Across the five unseen settings, GranFlow has the best average result on all six metrics in Table 6. Its MS-SSIM ranks second under extreme blur and noisy infrared input. Across the two local tests, it has the best average SCD, MSE, PSNR, and MS-SSIM, and the second-best QNCIE. Figure 11 shows less noise in the building region and clearer outlines in the blurred shop-sign region. Nearby scene content also remains visible.


W3/Q3 — Specific Innovations within VAE- and DiT-Based Frameworks

Each module in GranFlow serves a distinct and complementary role within a coordinated restoration and fusion process. Our contribution lies in how these roles are defined and connected, so that each part supports a different need in handling degraded inputs. This design links degradation-guided fusion with sample-specific latent refinement.

  • RFfusion adapts a VAE for fusion and prioritizes efficient one-step RF sampling. Its frequency and saliency losses encourage information preservation, but it does not explicitly learn degraded-to-clean latent transport. GranFlow trains a dual-branch VAE to restore and fuse degraded inputs, then learns conditional RF refinement between degraded and clean fused latents, directly targeting residual corruption.

  • DiTFuse uses VAE, DiT, and masked learning for instruction-controlled fusion and takes explicit text instructions as input. GranFlow requires neither text prompts nor degradation labels: image-derived wavelet cues and local latent interactions guide spatial-channel mixing, while current displacement supplies state information.

  • OmniFuse first restores each source latent with diffusion, then learns weights to combine the restored features. Its contrast, texture, and color losses also use the reconstructed source images as references. When source restoration leaves noise or distorts details, these errors can enter both the features being fused and the references used to train fusion. The method does not learn a further degraded-to-clean transformation of the fused latent to directly correct such remaining errors. GranFlow addresses this limitation by training a dual-branch VAE to restore and fuse the inputs, followed by conditional RF trained between degraded and clean fused latents.

The modules of GranFlow work together to select useful features and then correct the remaining errors. The wavelet branch describes the degradation characteristics of each modality. The dual-branch VAE uses these conditions to restore modality features, while the spatial-channel gate combines the conditions with local latent features to decide how much each modality contributes at each location and channel. This produces the fused latent that conditional RF further restores, with DiT predicting the updates. Within this refinement stage, scale-gated normalization controls residual feature scale, while the displacement-scale embedding supplies the magnitude of each sample's current latent change:

s_t=\operatorname{std}(\delta_t),\qquad r_t=\operatorname{MLP}\bigl(\log(1+s_t)\bigr).$$ Here, $z_L^f$ is the degraded starting latent and $z_t$ is the current latent. Their difference describes the change during restoration, and $s_t$ measures its size. Small residuals retain their original scale, while larger ones receive more normalization. The embedding $r_t$ separately supplies the size of the change, so the rectified flow network can use both its pattern and magnitude. A fixed summary of the degraded input supplies scene context throughout refinement, including the starting step when $s_t=0$. Together, these conditions let the flow network use the input context and the current sample's state when predicting the next update. Table 4 tests our design at the same or a similar parameter count: | Conditioning arrangement | SCD ↑ | MSE ↓ | PSNR ↑ | MS-SSIM ↑ | | ------------------------ | ----------:| ----------:| -----------:| ----------:| | Global gate | 1.2711 | 0.0482 | 14.8469 | 0.6754 | | Unified conditioning | 1.6737 | 0.0481 | 15.6486 | 0.6761 | | Timestep-scheduled scale | 1.8592 | 0.0348 | 15.9845 | 0.7281 | | GranFlow | **1.8639** | **0.0311** | **16.0177** | **0.7344** |
AuthorsOpenReview ↗

Response to Reviewer rs2e (Part.3)

Table 4 shows that the tested components serve their intended roles in restoration and fusion. The spatial-channel gate selects information locally. We designed the gate to adjust the contribution of each modality at different image locations. When the global-gate control forces these weights to be the same across locations within each channel, PSNR falls from 16.02 to 14.85 and MS-SSIM from 0.73 to 0.67. The two stages benefit from conditions matched to their tasks. The first stage uses modality-specific frequency information to guide restoration and fusion. The second stage uses the current latent change to guide further refinement. The unified-conditioning control merges the modality descriptors into one global condition and changes how both stages use the conditions. Compared to our approach, it has resulted in a certain decrease in performance. The displacement scale provides useful information about the current sample. We use this scale to tell the refinement network how much each sample’s latent has changed. The timestep-scheduled control replaces it with an average scale from training data, so every sample receives the same scale value at a given timestep.

Taken together, Table 4 shows that the local gate, task-specific conditions, and sample-specific displacement scale each serve a distinct role. Their combination helps GranFlow select useful modality features and refine the fused latent, leading to better fusion quality.


W4/Q4 — Direct Analysis of Gating and Displacement Conditioning

Figures 13–15 explain why we use both frequency information and latent displacement, and show how the gate adjusts the contribution of each modality across the image.

Figure 13: Different degradations cause different amounts of latent change. We compare clean and degraded versions of the same scene using the same frozen autoencoder. The curves show how much the fused latent changes as each degradation becomes stronger. Noise and blur show large early changes that diminish at higher strengths. Haze continues to produce larger changes over the tested range. Infrared stripes produce much smaller latent changes. These trends show that latent change depends on both the type and strength of degradation. A single degradation-type scale would miss these differences, thus causing the model to lose information.

Figure 14: Frequency changes and latent displacement provide different information. Each point relates the frequency content of the input degradation to the resulting latent change. The input degradation is measured by the difference between degraded and clean inputs, split into low- and high-frequency bands. For haze, larger low-frequency changes accompany larger latent displacement; for blur, high-frequency detail changes relate clearly to latent displacement. Noise behaves differently: its high-frequency changes cover a wide range, while latent displacement grows much less. Infrared stripes also show changes in input frequency content while the latent displacement remains small. These patterns explain why we use both kinds of information. Frequency features describe how the inputs are affected, while displacement scale measures how much the latent representation changes. The two signals therefore serve different roles: frequency information guides restoration and fusion of the source features, and the current displacement scale guides further refinement of the fused latent.

Figure 15: The gate adjusts modality weights by region. In our fusion rule, gg weights visible features and 1−g1-g weights infrared features. The last column compares the gate under degradation with the gate for the clean image pair. Red means that the visible weight increases; blue means that it decreases. Under visible noise, the gate lowers visible weights over much of the central background, while increasing them in some other regions. Under haze, visible weights increase around the pedestrians and decrease in several background regions. Under blur, the changes are mixed and follow different parts of the scene. Thus, the gate adjusts the visible–infrared balance locally according to both image content and degradation. It can retain useful visible features in one region while giving infrared features more weight in another. In the infrared-stripe example, the visible input is unchanged and the gate difference is close to neutral. The displayed modality weights therefore remain close to those for the clean pair. Across the examples, the gate shows different responses to different degradations and different regions. This illustrates its intended role in selecting and combining modality features locally.

Together, these analyses show how the parts of GranFlow work together to handle degraded inputs. The visual analyses and controlled experiments support the effectiveness of our method: its components serve different roles and work together to select useful information and correct the errors that remain after fusion.

Reviewer_K4b3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GranFlow is proposed for infrared-visible image fusion under different types of degradation, including noise, blur, stripe artifacts, and haze. The main idea is to use different conditioning mechanisms for frequency-related corruption and broader changes in image appearance, rather than treating all degradations in the same way. The framework has two stages. First, a wavelet-conditioned dual-branch VAE extracts features from the infrared and visible inputs and combines them using gated fusion. The fused latent is then refined by a conditional DiT through Rectified Flow. This second stage uses both a global summary of the initial latent and a scale embedding based on the current latent displacement. The authors evaluate GranFlow under single, compound, and degradation-free conditions on several datasets. The experiments include comparisons with existing fusion methods, component ablations, and additional object detection and semantic segmentation results.

Critical strengths

1.I find the motivation behind the wavelet-guided fusion reasonable. Noise, blur, and stripe artifacts affect the frequency content of images differently, and the two modalities may not suffer from the same degradation. Using wavelet information to adjust their contributions is therefore a sensible choice. The subsequent Rectified Flow stage also provides a way to refine the fused representation beyond the initial restoration and fusion result, without requiring degradation labels during inference. 2.The paper includes a useful range of experiments. Table 1 reports strong results on DeMMI-RF and EMS, although the advantage is not uniform across every metric. The authors also evaluate compound degradations and clean inputs, which helps assess whether the method remains useful across different conditions. I appreciate the inclusion of the component ablations in Table 4, the RF-Fusion comparison in Appendix A.7, and the downstream detection and segmentation experiments.

Critical weaknesses

1.My main concern is whether the experiments actually demonstrate the benefit of matching conditioning granularity to degradation characteristics. Table 4 includes individual module ablations and a timestep-only baseline, but these experiments do not directly show why the proposed matching of conditioning signals is better than other arrangements. A comparison with a capacity-matched unified conditioning scheme or a different assignment of the same conditioning signals would make this point clearer. The scale embedding is also computed from latent displacement, yet the paper does not directly show how this quantity relates to degradation severity or broader distributional changes. The results support the effectiveness of the complete model, but the specific contribution of the granularity-matching design remains unclear. 2.I am also concerned about generalization beyond the degradation patterns used in training. Appendix A.3 assigns noise and haze to visible images, stripe artifacts to infrared images, and blur to both modalities. These rules are understandable for constructing synthetic data, but real imaging conditions may not follow the same assignments. Although the paper includes real-world visual examples, more quantitative results on naturally degraded data or different corruption configurations would make the robustness claims stronger. Another issue is inference speed. Appendix A.10 reports only 0.11 FPS at 1280 × 1024 resolution with 25 sampling steps. This is quite slow for some of the time-sensitive applications discussed in the paper, and the benefit of the refinement stage needs to be considered alongside its computational cost.

Questions for authors

1.Could the authors compare the proposed granularity-matched conditioning with a unified conditioning scheme or other conditioning arrangements under similar model capacity? It would also help to clarify how the latent displacement scale relates to different degradation characteristics. 2.How does GranFlow perform when the modality-specific degradation assignments differ from those used in training, or when evaluated on independent data with naturally occurring degradations? 3.Given the reported 0.11 FPS at full resolution, could the authors provide a clearer comparison of the performance gains and computational overhead introduced by the Rectified Flow refinement stage?

4 comments
AuthorsOpenReview ↗

Response to Reviewer K4b3 (Part.1)

Thank you for your comments. We revised the manuscript; see Secs. 4.5–4.6 and Apps. 11–12, 15. Changes are marked in dark red.

W1/Q1 — Evidence for Granularity-Matched Conditioning

We added three capacity-comparable conditioning varients on DeMMI-RF (Table 4):

VariantQNCIE↑CC↑SCD↑MSE↓PSNR↑MS-SSIM↑
Global gate.8344.67861.2711.048214.8469.6754
Unified conditioning.8339.67811.6737.048115.6486.6761
Timestep-scheduled scale.8870.68851.8592.034815.9845.7281
GranFlow.8887.70301.8639.031116.0177.7344
  • Global gate. The global-gate varient spatially averages the fusion-gate logits before the sigmoid. This applies the same modality weight at every spatial position within a channel. It therefore constrains the fusion gate's ability to assign different visible/infrared contributions to different regions. GranFlow retains a spatial-channel gate generated jointly from modality-specific wavelet conditions and local latent interactions. The wavelet descriptors summarize each modality's frequency and statistical characteristics, while the local interactions allow the resulting weights to vary with image content at each position. This enables different modality contributions in regions with different information quality. GranFlow improves PSNR from 14.85 to 16.02 dB and MS-SSIM from 0.6754 to 0.7344 over the global-gate varient.

  • Unified conditioning. The unified-conditioning combines the two modalities’ wavelet descriptors into a shared global condition, uses global modulation and gating in the autoencoder, and jointly encodes timestep, latent summary, scale, and wavelet information in the rectified flow network. GranFlow instead keeps the wavelet conditions separate before fusion. Each condition adjusts its own modality’s features and helps compute fusion weights that vary across positions and channels. This lets information from each modality guide how much that modality contributes to each region. After fusion, the rectified flow network uses the timestep, a summary of the degraded fused latent, and a dynamic displacement-scale embedding to refine the fused features. At comparable model capacity, GranFlow improves all six metrics over this varient, including SCD from 1.67 to 1.86 and MS-SSIM from 0.6761 to 0.7344.

  • Timestep-scheduled scale. The timestep-scheduled varient keeps the same autoencoder and trains the rectified flow model separately. It retains the scale-embedding pathway but uses an average scale computed from the training data for each timestep. Thus, all samples at the same timestep receive the same scale. GranFlow instead computes the scale from the displacement between each sample’s current predicted latent and its degraded starting latent. After initialization, samples at the same timestep can therefore receive different scales. The scale embedding tells the velocity predictor how large the current displacement is. Compared with this varient, GranFlow reduces MSE from 0.0348 to 0.0311, increases MS-SSIM from 0.7281 to 0.7344 and improves PSNR from 15.9845 to 16.0177 dB.

The configurations follow the same data manifest, degradation pipeline, objectives, training lengths, and seed protocol (Line 1345). GranFlow outperforms these alternatives, supporting both spatially resolved modality selection and sample-dependent scale conditioning. Fig. 12(Line 1389) provides visual comparisons on a hazy scene. The global-gate and unified-conditioning variants lose much of the color. The timestep-scheduled variant retains color but has weaker contrast. GranFlow better preserves distinct colors and tonal contrast.

Relationship between displacement scale and degradation characteristics. Latent displacement measures how much degradation changes the fused latent representation. Figure 13(Line 1414) compares degraded and clean fused latents of the same scenes using the frozen autoencoder, with strength varied within each degradation type. Noise produces a rapid increase and then it tends to stabilize. The change amplitude of blur is smaller. Haze's displacement continues to increase with the change of degradation intensity. Stripe induce smaller displacements. RMS closely follows STD, while absolute mean shifts are much smaller. Different degradation types show different relationships between degradation strength and latent displacement. We therefore use each sample’s latent displacement to guide refinement. Figure 14(Line 1460) shows that the relationship between frequency residual energy and latent displacement varies across degradation types. They are positively associated in the low-frequency band for haze and in both bands for blur. For noise and stripes, residual energy varies, while latent displacement levels off for noise and remains small for stripes. These observations support our use of wavelet information to guide fusion and each sample’s displacement scale to guide refinement.

AuthorsOpenReview ↗

Response to Reviewer K4b3 (Part.2)

W2/Q2 — Generalization to Unseen and Real-World Degradations

We evaluate five unseen scenarios and two local degradations. They include stronger noise and blur, low light, overexposure, noise in the infrared image, and noise or blur limited to part of the image. Tables 18-21(Line 1299) give the detailed results, and Table 6 reports the averages. These tests examine whether GranFlow can still fuse useful information when the degradation strength, type, or location changes.

SettingSCDPSNRMS-SSIM
Extreme noise1.920815.30710.6618
Extreme blur1.835215.19850.6374
Low light1.936615.20920.6913
Overexposure1.732113.42900.6668
Noisy IR1.928415.38520.6788
Local noise1.904815.35940.6783
Local blur1.939315.48460.6972

The extreme noise and blur levels exceed the parameter ranges used for training. Noisy IR uses grayscale Gaussian noise. Averaging each metric over the five unseen scenarios yields Table 6, where GranFlow ranks first on all six metrics. Averaging the two local tests yields first place in SCD, MSE, PSNR, and MS-SSIM and second in QNCIE. These results support generalization across unseen strengths, exposure changes, and spatially restricted corruption. As is shown in Fig 11,under extreme noise, GranFlow reduces noise while keeping scene boundaries visible. Under extreme blur, structures are clearer than other methods. In the low-light example, the target and background are easier to see. Under overexposure, the bright area is less spread out, and nearby content remains visible. With noisy infrared input, pedestrian outlines remain clear without the excessive brightening seen in some competing results. Together with Tables 18-20, these examples show that GranFlow can preserve useful information under several unseen conditions, supporting the combined use of wavelet-guided fusion and refinement based on each sample's latent displacement. The last two rows of Fig. 11 show that GranFlow reduces local noise and blur while preserving surrounding content. Averaged across both tests, it ranks first in SCD, MSE, PSNR, and MS-SSIM, and second in QNCIE.

Table 5 and Fig. 5 also evaluate naturally degraded inputs. GranFlow obtains the best NIQE, second-best CLIP-IQA, and third-best FADE. These results demonstrate its effectiveness beyond synthetic degradations and support its ability to generalize to real-world scenes.


W2/Q3 — Fusion Quality and Computational Efficiency Refinement

GranFlow's computational cost acceptable for applications that prioritize fusion quality, given its strong results and competitive speed among the compared diffusion-based methods. Table 23 reports the highest SCD and PSNR among the listed methods, at 1.86 and 16.02 dB on DeMMI-RF. Figure 4 also shows improved haze removal and clearer scene structures. GranFlow achieves these results while running faster than DRMF, Text-DiFuse, and Mask-DiFuser, supporting the practical value of its restoration and fusion design.

The following FPS results are measured using the same hardware, batch size, and 1280×1024 resolution:

MethodFPS ↑
DRMF0.0219
OmniFuse0.1620
Text-DiFuse0.0140
Mask-DiFuser0.0065
GranFlow0.1140
GranFlow distilled0.6842

GranFlow is faster than three of the four baselines and also achieves strong visual results, as shown in Fig. 5.

To further reduce latency, we construct a smaller DiT and distill the original model into it. With the same 25 sampling steps, the distilled model runs approximately six times faster, with only minor decreases in fusion metrics (Table 22). It also reduces FLOPs by 91.3% and parameters by 77.0% (Table 23). These results show that GranFlow offers a favorable balance between fusion quality and computational cost, with distillation providing an efficient option for applications with tighter latency requirements.

We sincerely thank you for your thoughtful comments. We have revised the PDF and provided responses to the questions and concerns you raised. We hope these revisions and responses clarify the issues and weaknesses identified in your review, and we appreciate your constructive feedback.

Thank you for your careful review and constructive suggestions.

In response to your concerns, we have added comparisons of conditioning strategies, analyses of latent displacement, evaluations under unseen degradations, and results on computational cost and model distillation.

We would appreciate your feedback on whether these additions address your concerns and whether any points require further clarification.

Reviewer_vAJJOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes GranFlow for infrared-visible image fusion under degradations such as noise, blur, stripe, and haze. The author believes that different degradations require conditioning at different granularities: local selection of reliable modality information and sample-level control of latent refinement should be modeled separately. GranFlow consists of two stages. Stage I uses a dual-branch VAE with wavelet-derived conditioning, adaptive feature modulation, and spatial/channel gating to combine the two modalities. Stage II freezes the VAE and uses a conditional DiT with rectified flow to transform a degraded fused latent into a clean fused latent. The flow model receives timestep information, a global summary of the degraded latent, and a scale embedding derived from the displacement along the current trajectory. The evaluation covers individual degradations, compound degradations, degradation-free inputs, real-world scenes, object detection, and semantic segmentation.

Critical strengths

1.The problem is practically relevant, and the design is well motivated. Degradation changes the relative reliability of infrared and visible information. Fixed fusion strategies can retain corrupted content or discard useful complementary details. Combining frequency-aware modality selection with latent refinement provides a coherent approach to this problem.

  1. The evaluation is broad, and the main results are competitive. The single-degradation, compound-degradation and clean-input experiments provide sufficient support for the method across different conditions.
Critical weaknesses
  1. RFfusion[1] already combines a task-specific VAE, latent-space fusion, two-stage training, and rectified flow. Moreover, ControlFusion[2] includes a spatial-frequency collaborative adaptation mechanism, and TG-ECNet[3] proposed DeMMI-RF, which is the main dataset of this article. These methods are similar to the method in this article, but they have not been listed as comparison methods. The authors may consider adding these methods.

  2. This paper needs to consider data from other unseen scenarios besides non-degraded data. The degradation parameters of these degraded images are outside the range seen during model training and may even correspond to very extreme environments. The model’s performance on these images needs to be reported.

  3. The author can consider adding noise or blur to only one area of the visible image to keep the infrared clean, so as to prove that the model has the ability to deal with local degradation.

  4. The training data of the paper includes DVTOD, M3SVD, DroneVehicle, and LLVIP. The specific number and distribution of these data sets need to be clarified

[1] Wang Z, Zhang J, Guan T, et al. Efficient rectified flow for image fusion[J]. Advances in Neural Information Processing Systems, 2026, 38: 22927-22948. [2] Tang L, Wang Y, Cai Z, et al. Controlfusion: A controllable image fusion network with language-vision degradation prompts[J]. Advances in Neural Information Processing Systems, 2026, 38: 118304-118327. [3] Sun Y, Li X, Zhu P, et al. Task-gated multi-expert collaboration network for degraded multi-modal image fusion[C]//International Conference on Machine Learning. PMLR, 2025: 57571-57586.

Questions for authors

Questions can be found in the Critical Weaknesses. In addition, the authors are advised to revise the following: The text accompanying Table 11 claims the best CC on M3SVD, although GranFlow's 0.5851 is below Mask-DiFuser's 0.5872.

4 comments
AuthorsOpenReview ↗

Response to Reviewer vAJJ

Thank you for your valuable comments. We have carefully considered them and revised the manuscript to further demonstrate the effectiveness and soundness of our method. The relevant revisions can be found in App. 8 (p. 22, Line 1162), Sec. 4.5 (p. 9, Line 451), App. 11 (p. 24, Line 1262), and App. 3.1 (p. 16, Line 859). For the convenience of reading, the revised text is marked in dark red.


W1 — Comparisons with Related Restoration-Fusion Methods

We have included all three methods in Tables 15–16 (Line 1155). Compared with our method, each of them has architectural limitations. RFfusion uses a Rectified Flow model originally designed for natural image generation to finish fusion tasks, but it does not explicitly model degradation information. TG-ECNet uses expert modules with the same architecture for different degradation types, which limits their ability to adapt to the specific characteristics of each degradation. ControlFusion combines text and degradation information into a unified condition. Since this condition is a global signal, it cannot precisely determine which regions are degraded or adapt the restoration locally. Table 16 reports PSNR on DeMMI-RF and EMS:

MethodDeMMI-RFEMS
TG-ECNet14.616711.7935
ControlFusion13.509711.6678
GranFlow16.017713.4665

GranFlow ranks first on all six metrics among these three methods on both datasets. Table 15 further compares RFfusion under degradation-free and Haze+Blur conditions. Under Haze+Blur, RFfusion obtains PSNR/MS-SSIM of 14.1559/0.6301, compared with 15.2436/0.6793 for GranFlow. These metric results suggest that our Scale Embedding and Scale-Gated Residual Normalization adjust the restoration strength according to the magnitude of latent displacement in the rectified-flow process. This allows the model to adapt its refinement to individual samples and degradation severities, leading to better fusion quality.


W2 — Generalization to Unseen Degradation Settings

We have extended the evaluation to extreme noise at 80/25580/255 and extreme blur with kernel sizes ranging from 41 to 61. Table 18(Line 1299) reports PSNR/MS-SSIM of 15.3071/0.6618 for extreme noise and 15.1985/0.6374 for extreme blur. Tables 19-20(Line 1320) additionally evaluate low light, overexposure, and infrared grayscale Gaussian noise with a standard deviation of 25/25525/255. We average each metric over the five preceding unseen-scenario evaluations; the resulting aggregate is reported in Table 6(Line 461). GranFlow ranks first on all six averaged metrics, with SCD 1.8706, PSNR 14.9058, and MS-SSIM 0.6672.

Figure 11(Line 1269) provides qualitative comparisons for the five unseen scenarios. Across the Extreme Noise and Extreme Blur rows, GranFlow suppresses residual artifacts while preserving target contours and fine structures. Under low light and overexposure, it produces more balanced visibility and contrast; under infrared noise, it retains scene structure while using complementary information from the visible modality. These visual observations are consistent with the quantitative trends in Tables 18-20. The qualitative and quantitative results demonstrate that our proposed granularity matching mechanism can still distinguish between image information and degraded information in the face of untrained scenes, and complete the fusion task.


W3 — Robustness to Localized Degradations

Table 21(Line 1364) reports PSNR/SCD of 15.3594/1.9048 for local noise and 15.4846/1.9393 for local blur. Figure 11 separately visualizes these localized cases: GranFlow corrects the corrupted regions while keeping the surrounding image content comparatively stable, suggesting that the granularity matching mechanism is spatially targeted rather than globally over smoothing the image. This result greatly supports the effectiveness of our model.


W4 — Training Data Composition and Distribution

App.3.1(Line 859) now specifies the training-set composition: 1,790 pairs from DroneVehicle, 150 from DVTOD, 3,800 from M3SVD, and 4,000 from LLVIP, totaling 9,740 pairs. These correspond to 18.38%, 1.54%, 39.01%, and 41.07% of the prepared dataset, respectively. Images are standardized to 1280 x 1024, and training uses random 256 x 256 crops. In addition, we also introduced our degradation generation pipeline, which, together with our proposed granularity matching mechanism, greatly improves the generalization and robustness of our model.


Q1 — Correction of the M3SVD CC Ranking

We have revised the relevant text in the manuscript (Line 1112) and confirmed that this correction does not affect our conclusions.

We sincerely thank you for your thoughtful comments. We have revised the PDF and provided responses to the questions and concerns you raised. We hope these revisions and responses clarify the issues and weaknesses identified in your review, and we appreciate your constructive feedback.

Reviewer_vAJJOpenReview ↗

Response to the Authors' Rebuttal

The authors have addressed my previous concerns well by providing additional experiments and explanations, including comparisons with relevant reconstruction-based fusion methods, generalization to unseen scenarios, and robustness to local degradations. I also appreciate the corresponding revisions made to the main paper. These improvements have significantly enhanced the quality of the paper. We appreciate the authors' efforts. Since I have already assigned a relatively high score to this paper, I have decided to maintain my original rating.

AuthorsOpenReview ↗

Response to Reviewer vAJJ

Thank you for your constructive suggestions, which helped us better demonstrate GranFlow’s generalization to unseen scenarios and robustness to local degradations. We appreciate your positive assessment and the time you devoted to reviewing our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.