The workshop short paper “Beyond Raw Signals: Undecoded Generative Latents as Privileged Synthetic Data”, authored by Cristian Sbrolli, Nicolas Michel, Matteo Matteucci and Toshihiko Yamasaki and resulting from the collaboration between AIRLab at Politecnico di Milano and the CVM Lab at the University of Tokyo, received the Best Short Paper Award at the 3rd Workshop on Synthetic Data for Computer Vision (SynData4CV), held in Denver, United States, on June 4, 2026, held as part of CVPR 2026 – IEEE/CVF Conference on Computer Vision and Pattern Recognition.
The paper addresses the use of multimodal knowledge in computer vision scenarios where datasets provide only RGB images. Following the Learning Using Privileged Information paradigm, the proposed approach uses generative models to synthesize missing modalities and exploit them only during training.
Instead of decoding these synthetic modalities into raw signals, the method directly uses the undecoded generative latents as compact privileged information. A multimodal teacher model is trained with access to this additional information, and its knowledge is then transferred to an RGB-only student through the proposed MESSy loss. At inference time, the final model requires only RGB images, preserving standard deployment cost while benefiting from multimodal supervision acquired during training.