Research map / Image generation / Fast diffusion sampling
Efficient Rectified Flow for Image Fusion
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 RFfusion, a one-step, training-free-sampling diffusion image fusion method that combines Rectified Flow with a fusion-tailored latent VAE trained in two stages.
Problem
Diffusion-based image fusion requires hundreds of sampling steps and heavy computation, and VAE latent-space approaches are trained for reconstruction rather than fusion objectives.
Approach
Applies Rectified Flow to straighten the sampling path so that fusion-guided diffusion (DDFM-style posterior sampling with EM-derived fusion prior) runs in one step, operating in a task-specific VAE latent space whose decoder is fine-tuned with frequency and fusion losses in two stages.
Key results
- Runtime of 0.308 s on RoadScene versus 22.03 s for DDFM and 62.47 s for CCF, with 65.57M parameters.
- On M3FD, MI improves from 2.871 (DDFM) to 3.320 and VIF from 0.677 to 0.748.
- Ablation shows combining both VAE training stages yields PSNR 61.81 versus 59.41 without training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add both VAE training stages instead of using neither (ablation) | RFfusion · MEFB | PSNR: 59.41 -> 61.81 | Helped |
| add VAE stage I to stage II-only training (ablation) | RFfusion · MEFB | PSNR: 60.36 -> 61.81 | Helped |
| add VAE stage II to stage I-only training (ablation) | RFfusion · MEFB | PSNR: 59.68 -> 61.81 | Helped |
| add frequency similarity loss (ablation) | RFfusion · MEFB | PSNR: 57.22 -> 58.84 | Helped |
| add saliency mask loss (ablation) | RFfusion · MEFB | PSNR: 57.22 -> 59.67 | Helped |
| add both frequency similarity and saliency mask losses (ablation) | RFfusion · MEFB | PSNR: 57.22 -> 61.81 | Helped |
| replace DDFM sampling with RFfusion's one-step Rectified Flow and latent VAE approach | DDFM · M³FD | VIF: 0.677 -> 0.748 | Helped |
| replace DDFM sampling with RFfusion's one-step Rectified Flow and latent VAE approach | DDFM · T&R | SSIM: 0.207 -> 0.921 | Helped |
| use RFfusion instead of DDFM | DDFM · M³FD | MI: 2.871 -> 3.320 | Helped |
| use RFfusion instead of DDFM | DDFM · T&R | VIF: 0.277 -> 0.675 | Helped |
| use RFfusion instead of the fastest listed diffusion baseline | diffusion-based image fusion methods · RoadScene | Runtime (s): 1.997 -> 0.308 | Helped |
Surprising findings
- Using visible-light images as the starting input instead of pure Gaussian noise improves fusion performance, per the authors' observation.
Limitations
- Relies on a Rectified Flow model pretrained on generic image generation, not designed for fusion, which may limit fusion performance.
- Comparisons of efficiency and fusion metrics are reported on a single dataset (RoadScene) for runtime.
Experimental setup
Datasets: LLVIP MSRS
Benchmarks: M3FD TNO RoadScene MEFB MFIF (Lytro, MFFW, MFI-WHU)
Metrics: MI VIF SCD EN SSIM SF AG PSNR CC Qcb
Baselines: U2Fusion YDTR UMFusion ReCoNet LRRNet CoCoNet DDFM DRMF Dif-Fusion Diff-IF Text-DiFuse CCF
Builds on: DDFM Rectified Flow Latent Diffusion Models VQGAN
Closest papers by meaning
- Task-driven Image Fusion with Learnable Fusion Loss (CVPR 2025 · similarity 0.87)
- Taming Rectified Flow for Inversion and Editing (ICML 2025 · similarity 0.86)
- Improving the Training of Rectified Flows (NeurIPS 2024 · similarity 0.85)
- Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations (ICLR 2025 · similarity 0.85)
- Conditional Controllable Image Fusion (NeurIPS 2024 · similarity 0.85)
- Balanced Conic Rectified Flow (NeurIPS 2025 · similarity 0.85)
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.