Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Efficient Rectified Flow for Image Fusion
Authors: Zirui Wang, Jiayi Zhang, Tianwei Guan, Yuhan Zhou, Xingyuan Li, Minjing Dong, Jinyuan Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method outperforms other state-of-the-art methods in terms of both inference speed and fusion quality. Code is available at https://github.com/zirui0625/RFfusion. |
| Researcher Affiliation | Academia | Zirui Wang1, Jiayi Zhang2, Tianwei Guan3, Yuhan Zhou2, Xingyuan Li4, Minjing Dong1, Jinyuan Liu2 1 City University of Hong Kong 2 Dalian University of Technology 3 Chinese University of Hong Kong 4 Zhejiang University |
| Pseudocode | No | The paper includes figures illustrating the pipeline and mathematical formulations, but it does not contain any explicit pseudocode or algorithm blocks labeled as such. |
| Open Source Code | Yes | Code is available at https://github.com/zirui0625/RFfusion. |
| Open Datasets | Yes | For the infrared and visible image fusion task, evaluations are performed on three widelyused benchmark datasets: M3FD [1], TNO [38], and Road Scene [39]. For the multi-exposure and multi-focus fusion tasks, we utilize the MEFB [40] and MFIF [10] datasets, respectively. The MFIF dataset includes the Lytro [41], MFFW [42], and MFI-WHU [43] datasets. ... The two-stage training of the VAE was conducted entirely on an NVIDIA V100 GPU. In the first stage, the model was trained on the LLVIP [44] and MSRS [45] datasets for 20 epochs. |
| Dataset Splits | No | The paper mentions training for 20 and 40 epochs on LLVIP and MSRS datasets respectively, and evaluates on M3FD, TNO, Road Scene, MEFB, and MFIF datasets. However, it does not provide specific training/test/validation splits (e.g., percentages or counts) for any of these datasets within the main text. |
| Hardware Specification | Yes | The two-stage training of the VAE was conducted entirely on an NVIDIA V100 GPU. ... All experiments are conducted on an NVIDIA V100 GPU, and the fusion speed as well as the number of model parameters are evaluated on the Road Scene [39] dataset to comprehensively assess the efficiency and complexity of each method. |
| Software Dependencies | No | The paper describes the use of Variational Autoencoder (VAE) and Rectified Flow mechanisms, but it does not specify any software libraries or frameworks with their version numbers that were used for implementation (e.g., PyTorch 1.9, TensorFlow 2.x). |
| Experiment Setup | Yes | The two-stage training of the VAE was conducted entirely on an NVIDIA V100 GPU. In the first stage, the model was trained on the LLVIP [44] and MSRS [45] datasets for 20 epochs. Interestingly, the best validation performance was typically achieved within just 4 to 5 epochs. The second stage involved training exclusively on the MSRS [45] dataset for 40 epochs. The remaining hyperparameters for both stages were configured in accordance with the experimental settings detailed in [26] and [37]. |