Building Federated Multimodal AI Workflows with NVIDIA FLARE

Training across separate organizations can become expensive even when their raw data stays local: the model updates still have to travel. NVIDIA describes a smaller exchange in FedUMM's FLARE workflow. Each client keeps the base model fixed and trains small adapter components; the server combines those updates.

In an eight-client simulation, this reduced communication per client per round from 28.6 GB to 0.094 GB, while the reported visual-question-answering score improved by 0.7 points against full-model averaging. The result supports reducing update size in that experimental setting, which used public benchmarks and simulated sites.