When Vision Meets Graphs: A Survey on Graph Reasoning and Learning

This survey defines “vision meets graphs” as treating rendered graph images as first-class inputs rather than relying only on adjacency matrices, GNN message passing, or LLM graph serialization. Its central diagnostic is the Rendering–Perception–Inference framework: rendering determines which structural cues survive, perception maps pixels to nodes, edges, labels, and endpoints, and inference performs prediction or multi-step reasoning. Across graph reasoning, graph learning, and scientific graphs, the recurring bottleneck is perception: models often miss nodes or misbind edge endpoints before reasoning begins. Benchmarks such as VisionGraph cover 8 graph-theoretic task types, while VGCure spans 22 node-, edge-, and graph-level tasks; methods including Description-Program-Reasoning, GITA, and MCDGraph improve reliability by combining visual and textual inputs, explicit structural descriptions, augmentation, masked graph infilling, and contrastive discrimination. For learning, GVN and DEL augment GNNs with visual features, while GraphAbstract shows that natural-image-pretrained vision encoders can outperform GNNs on graph-level global-structure tasks, despite limitations on node- and edge-level prediction. Scientific graphs are especially promising because conventions reduce visual ambiguity: ChemDFM-X combines 5 chemical modalities, GIT-Mol aligns molecular images, graphs, and text, and domain models such as ChemVLM outperform general-purpose VLMs on chemical reasoning. ImageMol pretrains on millions of molecular images, whereas protein systems remain less mature: Geneverse uses LLaVA with AlphaFold-derived structure images, but LiveProteinBench finds that structural images can fail to improve, or even degrade, sequence-only prediction. The survey’s roadmap emphasizes graph-native visual pretraining, controllable rendering, active visual reasoning, and tool-based structural verification.