PPR OUTPUT
MULTIMODAL RETRIEVAL-AUGMENTED GENERATION
MG²-RAG
Multi-Granularity Graph for Multimodal Retrieval-Augmented Generation
A lightweight framework that preserves textual structure and fine-grained visual evidence in a unified multimodal graph, then retrieves across four semantic granularities with graph propagation.
43.3×average construction speedup
23.9×average cost reduction
Evaluated on retrieval, knowledge-based VQA, multimodal reasoning, and crisis classification.
Multi-Granularity Graph Case
Solid nodes form the PPR graph. Dashed sentence and object nodes visualize auxiliary retrieval pivots and their aggregation paths.
Drag nodes · scroll to zoom · drag canvas to pan
Graph Retrieval Example
Multimodal query signals activate the unified graph and rank the supporting chunk first.
Question
How many days does the gestation of this animal take?
How many days does the gestation of this animal take?Generated answer180 days
Recall@1 = 1.0
RETRIEVED EVIDENCE
“Gestation lasts six to seven months, following which a single calf is born and immediately concealed in cover.”
Qwen3.6 generation1804 completion tokens · finish: stop
Framework Overview
Construction aligns text entities with visual regions. Retrieval activates evidence from chunks, sentences, images and objects before propagating relevance over the unified graph.
Knowledge-Based VQA Cases
Retrieved multimodal evidence supports entity recognition, factual grounding and cross-modal alignment.
Construction Efficiency
Lightweight textual structure extraction and entity-driven visual grounding reduce construction time and monetary cost.
Citation
If you find MG²-RAG useful in your research, please cite our paper.
@inproceedings{dai2026mg2rag,
title={MG$^2$-RAG: Multi-Granularity Graph for Multimodal Retrieval-Augmented Generation},
author={Sijun Dai and Qiang Huang and Xiaoxing You and Jun Yu},
booktitle={Proceedings of the 19th European Conference on Computer Vision (ECCV)},
year={2026},
}