Research map / Image generation / Image restoration
ReFusion: Learning Image Fusion from Reconstruction with Learnable Loss Via Meta-Learning
Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, Shuang Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add learnable fusion loss weights w and v instead of fixed 1/2 (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 37.49 -> 42.89 | Improved |
| add meta-learning training of the loss proposal module (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 38.96 -> 42.89 | Improved |
| add separate fusion and reconstruction updates (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 39.33 -> 42.89 | Improved |
| use max rather than sum in the reconstruction loss (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 32.65 -> 42.89 | Improved |
| use learned feature fusion network instead of decision-based weighted sum (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 38.42 -> 42.89 | Improved |
| use attention fusion module instead of concatenation (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 39.26 -> 42.89 | Improved |
| add interactive feature extraction (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 39.25 -> 42.89 | Improved |
| add gating feature refinement (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 39.53 -> 42.89 | Improved |
| increase alpha1 from 1 to 10 in fusion loss (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 42.89 -> 41.04 | Hurt |
| vary alpha2 between 0.1 and 10 (ablation) | ReFusion · Infrared-visible fusion, MSRS | SD: 42.72 to 42.89 | No clear effect |
| use ReFusion instead of CDDFuse | CDDFuse · Infrared-visible fusion, MSRS | SD: 43.38 vs 42.89 | Mixed |
| use ReFusion instead of DeFusion | DeFusion · Medical image fusion, Harvard | SSIM: 0.76 -> 0.75 | Hurt |
Closest papers by meaning
- Efficient Rectified Flow for Image Fusion (NeurIPS 2025 · similarity 0.88)
- FreeFusion: Infrared and Visible Image Fusion via Cross Reconstruction Learning (TPAMI 2025 · similarity 0.88)
- Task-driven Image Fusion with Learnable Fusion Loss (CVPR 2025 · similarity 0.87)
- Test-Time Dynamic Image Fusion (NeurIPS 2024 · similarity 0.85)
- Learning Efficient Fuse-and-Refine for Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.85)
- Balanced Conic Rectified Flow (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
