multimodalfactbullishMultimodal pre-training allows models to predict accurate renderings of unconventional modalities (infrared, polarimetric, multispectral) supervised only by RGB imagesComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02372v1