The images don't need to be aligned. They don't even need to be of the same scene.
Infrared and visible image fusion requires training data where the two modalities are precisely aligned — same scene, same moment, pixel-registered. Getting this data is expensive: the cameras must be rigidly mounted, calibrated, and synchronized. The alignment process is labor-intensive (arXiv:2603.21820).
The alternative: train on arbitrarily paired data. Take any infrared image and any visible image — not necessarily of the same scene — and train the fusion model on these unpaired inputs. The model learns cross-modal relationships from the statistical structure of each modality rather than from pixel-level correspondence.
The result: comparable performance to models trained on datasets 100× larger using traditional strict pairing. The data efficiency gain is enormous — not because the model needs less data, but because it accepts data that was previously unusable (unaligned, unpaired, mismatched).
Validated across CNN, Transformer, and GAN architectures. The finding is architecture-agnostic — the data paradigm change helps regardless of the model.
The structural insight: the strict pairing requirement assumed that the model needed to learn which features correspond between modalities. The arbitrary pairing shows that the model mostly needs to learn what each modality contains (infrared: thermal structure; visible: texture and color) and how to combine them. The correspondence isn't learned from paired examples — it's inferred from the statistical properties of each modality independently. The pairing was never the learning signal. The modality structure was.