MultiBench

Why we included this project

Comparing multimodal models is usually a mess of incompatible setups, and MultiBench exists to fix that. It bundles 15 datasets spanning affective computing, healthcare, robotics, finance, HCI, and multimedia, and evaluates models on more than accuracy: it also tracks training and inference cost and how well a model holds up when modalities are noisy or missing. The companion MultiZoo toolkit ships roughly 20 fusion methods, objective functions, and training structures in modular form, so you can swap components instead of reimplementing every baseline. For researchers and teams that want reproducible comparisons, that combination is genuinely useful, and the documented process for adding new datasets and algorithms makes it easy to extend the benchmark to your own problem.