Research map / Label-efficient and robust learning / Test-time adaptation

Test-Time Dynamic Image Fusion

NeurIPS 2024 · Label-efficient and robust learning › Test-time adaptation · approach: Test-time adaptation of vision-language models · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Test-Time Dynamic (TTD) image fusion, which weights each source by a pixel-wise Relative Dominability derived from its uni-source reconstruction loss, with a generalization-bound argument, and applies it training-free to existing fusion models across VIF, MIF, MEF and MFF.

Problem

Existing dynamic image fusion methods rely on heuristic designs and lack theoretical guarantees or interpretability, risking unstable fusion results; the paper asks whether dynamic image fusion can be justified theoretically.

Approach

Decompose the fused image into uni-source components and show the generalization error bound shrinks when fusion weights are negatively covariant with each source's reconstruction loss; instantiate this as pixel-wise Relative Dominability, a softmax of negative losses, used as a test-time fusion weight with no training.

Key results

• Applying TTD to IFCNN on LLVIP raises SD from 37.75 to 38.99 and SF from 18.18 to 19.40 without any training.
• On MRI-PET, PIAFusion+TTD improves SD from 73.32 to 75.79 and SF from 30.41 to 31.65 over PIAFusion.
• Loss-positively-correlated weights degrade performance below static fusion, while the negatively correlated RD weights improve it, per the paper's ablation.

Measured effects reported in the paper

Surprising findings

• Positively correlated loss-based weights perform worse than static fusion, supporting the claimed sign requirement.

Limitations

• The generalization argument assumes a well-trained fusion model and a covariance condition that is not verified in general.
• Loss-based RD is pixel-level only for losses that give pixel maps; other losses need a hand-selected gradient channel chosen empirically.

Experimental setup

Datasets: LLVIP MSRS Harvard medical image dataset MEFB MFI-WHU

Benchmarks: LLVIP MSRS Harvard medical image dataset MEFB MFI-WHU

Metrics: EN SD AG EI SF SCD CE SSIM

Baselines: DenseFuse CDDFuse U2Fusion DDFM DeFusion PIAFusion MUFusion IFCNN SwinFuse TC-MoA

Builds on: IFCNN CDDFuse PIAFusion

Closest papers by meaning

• Conditional Controllable Image Fusion (NeurIPS 2024 · similarity 0.84)
• Hybrid-Supervised Dual-Search: Leveraging Automatic Learning for Loss-Free Multi-Exposure Image Fusion (AAAI 2024 · similarity 0.84)
• Predictive Dynamic Fusion (ICML 2024 · similarity 0.84)
• Revisiting Generative Infrared and Visible Image Fusion Based on Human Cognitive Laws (NeurIPS 2025 · similarity 0.83)
• Efficient Rectified Flow for Image Fusion (NeurIPS 2025 · similarity 0.83)
• MoDGS: Dynamic Gaussian Splatting from Casually-captured Monocular Videos with Depth Priors (ICLR 2025 · similarity 0.83)