inference-aware-learning

iFAN is a training framework that improves mask transformers for segmentation by aligning query ranking with mask quality and distilling intermediate predictions to the final layer, yielding consistent gains across benchmarks.