Image-Fusion is a reading list for deep learning image fusion, with a metrics folder attached

Two directories, one of which is a bibliography

The top level of the repository contains only two things, the README and a directory called `General Evaluation Metric/`. That single fact sets the scope. This is not a library and not a framework, and there is no API to call and no way to install it. It is the reference list for a survey of deep learning based image fusion, written as a running news feed of papers with the publication venue, a link to the paper and, for most entries, a link to a separate code repository. The primary language recorded for the project is MATLAB, which fits the one code folder, since evaluation metrics for image fusion are traditionally written as short scripts that produce a handful of numbers per image pair. The README also points at the author's personal homepage for the complete paper list and the latest research news, which is the place to go when you want the wider body of work rather than this one topic.

The survey is the anchor, and it has been recognised twice

The survey, a review of image fusion methods based on deep learning, was accepted by the Journal of Image and Graphics on 2022-07-29, and it is the item that the rest of the repository hangs from. Its path through the awards is worth following because it tells you the survey is the group's own work rather than a third party compilation. In September 2024 it was selected as a high impact paper by the Journal of Information Aerospace Technology, which the announcement says was one of only fourteen in the country that year, and in December 2024 it received the journal's outstanding paper award for 2024. Then on 2025-11-26 the same survey was named an outstanding paper for the 2020 to 2024 period. So the reading list has a paper behind it that has been through peer review, which is a different status from a hand maintained list of links, and the link to the survey's code is the same repository you are reading.

Recent work leans on degradation priors and diffusion models

The 2024 to 2026 entries show where the field has moved, and the direction is consistent. DRMF, published at ACM Multimedia 2024, is a degradation robust multi-modal image fusion method built on a composable diffusion prior. Mask-DiFuser, accepted by IEEE Transactions on Pattern Analysis and Machine Intelligence in September 2025, is a masked diffusion model for unified unsupervised image fusion. ControlFusion, accepted at NeurIPS 2025, adds controllability through language and vision degradation prompts. And DSPFusion, accepted by IEEE Transactions on Image Processing on 2026-06-02, does image fusion via degradation and semantic dual-prior guidance. The thread running through them is the idea that fusion is hard because the two input modalities are degraded differently, and that a prior describing the degradation is a better guide than a hand designed loss. For anyone choosing a starting point, the diffusion based work is where the current papers are, and the repository lists them in date order with links to both the paper and the code.

A long argument about whether fusion helps high-level vision tasks

A different thread in the list is the argument about what fusion is for. In June 2023 the group published a paper in Information Fusion titled on rethinking the necessity of image fusion in high-level vision tasks, proposing a practical infrared and visible image fusion network based on progressive semantic injection and scene fidelity. SeAFusion, a semantic-aware real-time infrared and visible image fusion network, followed and won an Information Fusion best paper award for 2024, announced in November 2024. C2RF, accepted by the International Journal of Computer Vision in March 2025, bridges multi-modal image registration and fusion through commonality mining and contrastive learning, which is a different problem again since it handles the two inputs that are not perfectly aligned. VideoFusion, accepted at CVPR 2026, extends the idea to video with a spatio-temporal collaborative network for multi-modal video fusion. Read together, these entries show a field that spent years on the pixel level question of what a good fused image looks like and then moved to asking what the fused image is for.

The table of contents is a taxonomy of the field

The README's table of contents is organised by application, and it is the fastest way to understand the scope of the survey. Multi-modal image fusion is the largest branch, split into infrared and visible image fusion and medical image fusion. Digital photography image fusion covers multi-exposure and multi-focus fusion, the two cases where a single scene is captured several times or at different focal depths. Remote sensing image fusion covers pansharpening, which is the merging of a high resolution panchromatic band with lower resolution multispectral bands. Then there is a section for a general image fusion framework, which is the attempt to write one architecture that handles more than one of these, and separate sections for the survey itself, for datasets and for evaluation metrics. The dataset section points at M3SVD, a large scale multi-modal multi-scene video dataset for infrared and visible video fusion, released in February 2025, which is the kind of resource that changes what a paper can claim.

The evaluation metric folder is the only executable part

Everything above is prose and links, and the one directory that holds code is `General Evaluation Metric/`, with the README's own contents listing a section for general evaluation metrics and a separate one for remote sensing image fusion metrics. That is the part of the repository you can run, and it matters more than a casual reading of a paper list would suggest, because image fusion has no single agreed measure. A fused image can be evaluated on how much information from each input it retains, on how sharp it is, and on how much artefact the fusion introduced, and different papers pick different subsets, which is the underlying reason results in this field are hard to compare across papers. The general metric folder is where a consistent set of numbers can be produced for a given pair of images, and the remote sensing section covers the pansharpening case, where the conventions differ again because the reference data is usually simulated from a lower resolution band. The README does not describe the files inside the folder, so the practical first step for a user is to look there.

Citations to copy, and the licensing question left open

The README asks readers to cite the survey if the summary helps them, and it supplies three BibTeX entries to copy, for Mask-DiFuser in IEEE TPAMI, for ControlFusion in NeurIPS, and for C2RF in the International Journal of Computer Vision. That is a small but useful detail, since hand typed citations in this field are easy to get subtly wrong, and the entries include the author lists and the venues as the group published them. The repository states no licence, and the top level listing contains no licence file, so there is no permission granted for redistribution of the list or the metric code, which is worth raising if you plan to reuse either. There are no published releases either, and the last push to the repository was on 2026-08-05, so the content is a moving document rather than a versioned artefact. The two things to rely on are the survey paper, which is a citable publication, and the code folders of the individual papers, which have their own terms.

Editorial conclusion

Image-Fusion suits a researcher who needs a map of the deep learning image fusion field and a maintained pointer to the code behind each paper, and it suits a student looking for the survey in one place with a worked set of evaluation metrics. It is a poor fit if you want a library you can call, since the repository is a bibliography with one code folder rather than a package, and it is a poor fit if you need a clear licence, since none is stated. Before you build on it, check three things: that the evaluation metric folder matches the metrics you plan to report, since the repository is the only executable part and its contents are not described in the readme, which metric a given paper used, because the field has several and the survey is where the comparison lives, and the state of the repository, with the last push on 2026-08-05 and no published releases.

Frequently asked questions

Official sources

1. Issues
2. Linfeng-Tang/Image-Fusion on GitHub
3. README