DistilVDR: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation
A 524M-parameter retriever keeps most of an 8B visual-document teacher’s accuracy while cutting index size and indexing cost.
DistilVDR is trained from a frozen vision-language teacher’s embedding space, without relevance labels, negative sampling, or a contrastive objective. Its asymmetric design gives more capacity to image-document encoding while keeping the text-query side at 70M parameters. The HiRes variant reaches 61.74 average NDCG@5 across ViDoRe v1–v3, reported as 86.9% of the 8B teacher. Both released variants store one million documents in a 15.6x smaller index than the strongest reproduced sub-1B multi-vector baseline and index the corpus about an order of magnitude faster. HF Daily Papers' note
DistilVDR is trained from a frozen vision-language teacher’s embedding space, without relevance labels, negative sampling, or a contrastive objective. Its asymmetric design gives more capacity to image-document encoding while keeping the text-query side at 70M parameters. The HiRes variant reaches 61.74 average NDCG@5 across ViDoRe v1–v3, reported as 86.9% of the 8B teacher. Both released variants store one million documents in a 15.6x smaller index than the strongest reproduced sub-1B multi-vector baseline and index the corpus about an order of magnitude faster. HF Daily Papers' note
score 5