LPIFM: Learned Perceptual Image Fusion Measure: model weights, preference labels, and reproducibility files

LPIFM (Learned Perceptual Image Fusion Measure) is a source-conditioned pairwise preference model for infrared–visible image fusion (IVIF). Given IR and VI sources and two fused candidates, it predicts A better / B better / Tie, and supports tie-aware Bradley–Terry (T-BT) ranking of method pools. This Zenodo record is the companion data and weights archive for the paper “Ranking Image Fusion the Way Humans Do: A Learned Pairwise Preference Measure for Infrared–Visible Fusion Assessment”. Contents: weights/lpifm_vifb_baseline_v1.pt — main VIFB-trained checkpoint (about 2 GB full training dump; inference uses SWA by default) weights/lpifm_evafusion_ft_v1.pt — optional EVAFusion preference fine-tune (paper: best checkpoint at epoch 3) preference_dataset_vifb/ — LPIFM-authored VIFB pairwise preference labels and protocol splits evafusion_finetune_dataset/ — LPIFM preference splits and reconstruction metadata (no EVAFusion images) configs/, objective_metrics/Objective_Metrics.xlsx, licenses, and CHECKSUMS.sha256 Not included: third-party VIFB or EVAFusion images. Obtain VIFB from https://github.com/xingchenzhang/VIFB and EVAFusion from its original authors. Code: https://github.com/HaoranLiu507/LPIFM (AGPL-3.0) Weights mirror: https://huggingface.co/FengShaner/LPIFM Licenses: LPIFM weights and LPIFM-authored labels — CC BY-NC-SA 4.0; companion source code on GitHub — AGPL-3.0.