Research map / Graph learning / GNN architectures
The Underappreciated Power of Vision Models for Graph Structural Understanding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Renders graphs as images with standard layout algorithms and applies vision encoders (ResNet, ViT, Swin, ConvNeXtV2) to graph tasks, and introduces GraphAbstract, a scale-shifted benchmark of four structural tasks where vision models generalize better than GNNs.
Problem
GNNs use bottom-up message passing, and existing graph benchmarks couple node features with topology, so they do not isolate whether models perceive global structure as humans do.
Approach
Translate graph topology into images via layout algorithms and apply off-the-shelf vision backbones without graph-specific modifications, then evaluate on a new benchmark (GraphAbstract) with feature-free topology tasks and graph-size OOD shifts.
Key results
- On topology classification, vision models drop only about 5-6% accuracy from ID to Far-OOD, while basic GNNs with one-hot degree features drop over 45%.
- On symmetry classification, vision models with spectral layouts reach about 20% higher accuracy than the best GNN variants, per the paper's claim.
- On the graph classification benchmarks, vision models are competitive with GNNs (e.g., ConvNeXt best on IMDB-MULTI at 53.5% and ENZYMES at 41.7%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use vision models instead of GNNs on GraphAbstract symmetry classification | GraphAbstract models · Symmetry Far-OOD | accuracy: 68.63 ± 0.97 -> 85.67 ± 1.06 | Helped |
| Use vision models instead of GNNs on GraphAbstract topology classification | GraphAbstract models · Topology Far-OOD | accuracy: 85.33 ± 9.94 -> 90.33 ± 4.60 | Helped |
| Use vision models instead of GNNs on GraphAbstract spectral gap regression | GraphAbstract models · Spectral Gap Far-OOD | MAE: 0.0571 ± 0.0154 -> 0.0946 ± 0.0094 | Hurt |
| Use vision models instead of GNNs on GraphAbstract bridge count regression | GraphAbstract models · Bridge Count Far-OOD | MAE: 4.1380 ± 0.2873 -> 3.6814 ± 0.1217 | Helped |
| Add LapPE to GCN+Degree (ablation) | GCN on GraphAbstract · Topology Far-OOD | accuracy: 33.67 ± 3.56 -> 55.00 ± 3.08 | Helped |
| Add SignNet to GCN+Degree (ablation) | GCN on GraphAbstract · Topology Far-OOD | accuracy: 33.67 ± 3.56 -> 77.93 ± 2.77 | Helped |
| Add SPE to GCN+Degree (ablation) | GCN on GraphAbstract · Topology Far-OOD | accuracy: 33.67 ± 3.56 -> 72.33 ± 7.90 | Helped |
| Use a spectral rather than force-directed graph layout (ablation) | Vision models · Symmetry ID | accuracy: 80.93 ± 1.09 -> 92.50 ± 0.43 | Helped |
| Scale GPS+SPE parameter count to match or exceed vision model capacity (ablation) | GPS+SPE · GraphAbstract | performance | Hurt |
Surprising findings
- Vision models memorize training data strongly yet show large generalization gaps on the standard benchmarks.
- Scaling GPS+SPE parameters to match vision model capacity degraded performance rather than improving it.
Limitations
- No direct comparison with human performance on the GraphAbstract tasks.
- Lacks theoretical characterization of why particular layouts help learning.
- Vision models show substantial generalization gaps on standard benchmarks and higher compute cost.
Experimental setup
Datasets: NCI1 IMDB-MULTI ENZYMES IMDB-BINARY PROTEINS MUTAG Cora
Benchmarks: GraphAbstract NCI1 IMDB-MULTI ENZYMES IMDB-BINARY PROTEINS
Metrics: accuracy mean absolute error
Baselines: GCN GIN GAT GPS LapPE SignNet SPE
Builds on: GIN GPS SignNet SPE
Closest papers by meaning
- Contextual Structure Knowledge Transfer for Graph Neural Networks (AAAI 2025 · similarity 0.89)
- Weisfeiler and Lehman Go Paths: Learning Topological Features via Path Complexes (AAAI 2024 · similarity 0.89)
- Open Your Eyes: Vision Enhances Message Passing Neural Networks in Link Prediction (ICML 2025 · similarity 0.89)
- GITA: Graph to Visual and Textual Integration for Vision-Language Graph Reasoning (NeurIPS 2024 · similarity 0.88)
- VisionGraph: Leveraging Large Multimodal Models for Graph Theory Problems in Visual Context (ICML 2024 · similarity 0.88)
- Vision Graph Prompting via Semantic Low-Rank Decomposition (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.