acceptodds
Under review as a conference paper at ICLR 2027

TeXIR: Reliability- and Integrability-Aware Structure Image Fusion

Abstract

Visible and infrared image fusion involves two fundamental decisions: what information to preserve and how that information should appear in the output. Most existing methods map the two modalities directly to a fused image, coupling information selection with image formation. We introduce TeXIR, which breaks this coupling by first constructing an explicit TeX-inspired scene state. TeXIR builds this state according to state-specific Reliability and, for structure, enforces Integrability so that locally reliable contributions form a globally coherent representation. Once the scene state is fixed, a separate readout controls its visual presentation without revisiting the fusion decision. The same fused state can therefore support different appearance conditions while preserving the underlying scene content. Experiments on FMB, MSRS, and RoadScene show that TeXIR preserves complementary source information while achieving strong perceptual quality across diverse scenes. Different readouts further exhibit the intended separation between source fidelity and visual presentation from the same fused state. TeXIR thus formulates visible and infrared image fusion as reliable scene state construction followed by controllable image formation. Code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image fusion

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wg7jOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TeXIR, a visible–infrared image-fusion framework that separates scene-state construction from image formation. It defines operational states for structure, region identity/color, and thermal response; estimates state-specific reliability to weight local source contributions; applies an integrability-aware optimization to directional structure preferences before screened-Poisson reconstruction; and uses separate readouts, including a frozen image-editing model, to produce different visual appearances from the same fused state. Experiments on FMB, MSRS, and RoadScene compare against 14 fusion methods using conventional fusion metrics and no-reference perceptual metrics. Additional analyses examine semantic segmentation, runtime, reliability maps, integrability, and parameter sensitivity. The central claim is that explicit state construction can preserve complementary scene information while allowing appearance to be changed independently at readout time.

Critical strengths
  1. A clear separation between information fusion and visual presentation. The paper introduces a TeX-inspired state representation that separates structural information, region and color information, and thermal responses. By constructing a fused state before rendering the final image, the framework provides a coherent approach to separating information selection from visual appearance.
  2. Explicit modeling of reliability and structural integrability. The method distinguishes local modality reliability from global structural consistency. State-specific reliability guides the contribution of visible and infrared information, while integrability-aware reconstruction reconciles local structural preferences into a coherent result. The experiments comparing direct fusion with integrability-aware reconstruction provide supporting evidence for this design.
Critical weaknesses

Some indicators and pictures in the article seem to be wrong. Please refer to the author's question for details.

Questions for authors

1.Figure 8 reports that TeXIR has 5.934 M parameters, yet the method relies on frozen pre‑trained components including SAM3 and Qwen‑Image‑Edit. The paper does not clarify whether these components are included in the reported parameter count, nor how their computational costs are handled in runtime evaluations. Although the efficiency section evaluates the deterministic (Xcv) readout output, the parameter‑counting protocol remains ambiguous, especially regarding the frozen region‑segmentation model. The authors should explicitly report the parameter count of trainable components, the total number of parameters required for each inference setup, as well as corresponding runtime, with clear specification of details for both the state‑construction and readout stages. Otherwise, the reported model size and efficiency may fail to reflect the full computational cost of the proposed framework. 2.This paper targets the task of image fusion, yet it leverages image‑enhancement‑oriented techniques within its pipeline. This creates a deviation from the core fusion objective. Since the competing baseline methods do not adopt such image‑enhancement components, the resulting comparison is unfair and unreasonable. 3.In Figure 2, several heat‑map visualizations look pairwise identical. It is unclear whether this stems from visualization artifacts, plotting bugs, or actual duplicated content in reliability observations. The authors need to verify the underlying raw data, check plotting code, and provide clarification. If genuine duplication exists among different reliability maps, this would indicate a flaw in state‑specific reliability estimation and undermine one of the core designs of TeXIR.

Reviewer_8dsnOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents TeXIR, a visible and infrared image fusion framework that separates scene state construction from image formation. TeXIR constructs a TeX inspired fused state using state specific reliability and integrability aware structure reconstruction, and then reuses this state to produce multiple deterministic and natural image readouts. The method is evaluated on FMB, MSRS, and RoadScene, together with additional studies on semantic segmentation, generalization, and ablations.

Critical strengths
  1. The paper presents a clear and conceptually interesting formulation of visible and infrared image fusion. Separating what information should be preserved from how it should be presented is well motivated, and the reusable fused state provides a natural way to support different output requirements without repeating the fusion process.
  2. The overall method is technically coherent. Reliability determines local modality contributions, while integrability reconciles these local structural decisions before reconstruction. The experimental evaluation is also comprehensive, covering three benchmarks, perceptual quality, semantic segmentation, generalization, ablations, and targeted analyses of the main components.
  3. The paper is generally well written and organized. The roles of the main components are clear, and the appendix provides substantial implementation details and additional analysis.
Critical weaknesses
  1. A central claim is that different natural image readouts can change appearance while preserving the same underlying scene content. The current evidence for this claim is mainly qualitative. Since these outputs rely on a generative image editing model, a direct quantitative evaluation of content preservation, especially for small objects and fine boundaries, would strengthen the paper.
  2. The semantic segmentation results provide useful downstream validation, but the contribution of individual design choices is not fully isolated. In particular, it would be useful to better understand the effects of integrability refinement and the visual form presented to the segmentation model.
  3. The convergence behavior of the optimization based structure reconstruction is not analyzed in much detail. The paper specifies iteration limits and stopping criteria, but the practical tradeoff between optimization accuracy and computational cost remains unclear.
Questions for authors
  1. Can the authors provide a quantitative characterization of content consistency between the observation and daylight readouts? Object correspondence, boundary consistency, or feature level matching would complement the qualitative examples, particularly for small or low visibility objects.
  2. The integrability analysis shows improved structural consistency after refining the directional preferences. Does this refinement also lead to measurable improvements in semantic segmentation? An ablation that removes only the circulation term while keeping the remaining reconstruction pipeline fixed would help establish this connection.
  3. The TeX readout assigns fixed hues to region indices, although these hues do not encode semantic classes, material identities, or natural object colors. How sensitive is the semantic segmentation result to this mapping? Testing several hue permutations or palettes while keeping the fused state fixed would help clarify this point.
  4. How quickly do the directional preference optimization and screened Poisson reconstruction converge in practice? It would be useful to know whether earlier stopping can retain similar reconstruction quality while reducing computation.
Reviewer_8nQmOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a visible and infrared image fusion framework that separates fused state construction from visual presentation. Specifically, the proposed method uses state specific reliability to determine local modality contributions and introduces integrability aware structure reconstruction before screened Poisson reconstruction. The obtained fused state organizes structural, region, chroma, and thermal information and supports several deterministic and natural image readouts. Extensive experiments show the advantages of the proposed method.

Critical strengths
  • This paper presents a clear and well motivated formulation of visible and infrared image fusion. Separating what information should be preserved from how it should be presented addresses an important limitation of image-level fusion. The explicit fused state and separate readout provide a natural way to make this distinction operational.

  • The distinction between reliability and integrability is made clear. Reliability determines local modality contributions, whereas integrability reconciles these local decisions into a coherent structure. The controlled analysis of the circulation term supports this design.

  • This paper compares against a broad set of recent methods on three benchmarks. The multiple readouts also demonstrate useful flexibility under different output requirements.

Critical weaknesses
  • The downstream experiments do not fully show how much each design contributes to semantic segmentation. In particular, the roles of reliability, integrability, and the reconstruction anchor are still unclear. It would be better to extend the ablation studies to segmentation using the same readout and frozen CLIP model, so that their individual contributions can be better understood.

  • The authors use XvX_v as the reconstruction anchor in Eq. (6), but the reason for this choice is not sufficiently explained. Since visible images can be severely degraded under low-light conditions, would a reliability-adaptive or infrared-based anchor be more appropriate, especially for nighttime scenes?

  • Table 1 only includes conventional fusion metrics for XCvXC_v, XCfXC_f, and the daylight readout, leaving the other two output forms less explored. A more complete comparison across all five readouts would help readers understand their respective strengths and tradeoffs. For the TeX output, the authors should also clarify which metrics are appropriate given its different color encoding.

  • Although preservation constraints are introduced, the generative readouts may still miss small targets or produce content that is not supported by the original observations. These potential failure cases deserve more attention, as they could affect the reliability of downstream tasks. Some quantitative analysis of target omission and unsupported generation would make the evaluation more convincing.

Questions for authors

Please see the weaknesses section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.