What Drives Compositional Generalization? The Importance of Continuous Training Objectives in Visual Generative Models
Researchers have investigated the factors influencing compositional generalization in visual generative models, focusing on how novel combinations of known concepts are generated. Their study highlights the significance of whether the training objective uses a discrete or continuous distribution, and the amount of information provided by conditioning during training. The findings suggest that incorporating a continuous, JEPA-based objective alongside a discrete loss, such as in MaskGIT, can enhance compositional performance in existing discrete models. AI
IMPACT Identifies key training objective characteristics that improve novel concept combination in visual generative models.