Research map / Graph learning / Graph self-supervised learning
Patch-Wise Graph Contrastive Learning for Image Translation
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.
Adds a shared-adjacency patch graph (TAGConv GNN with top-K graph pooling) and infoNCE node matching to CUT-style one-sided unpaired image translation, reporting state-of-the-art FID/KID.
Problem
Patch-wise contrastive learning for unpaired image-to-image translation uses point-wise patch matching and ignores topology/neighbor relations among patches, limiting high-level semantic correspondence between input and translated output.
Approach
Build a patch graph from cosine similarity of pretrained VGG16 features on the input, share its adjacency matrix with the output graph, apply topology adaptive graph convolution to node features, maximize node-wise mutual information via infoNCE, and add top-K graph pooling as a learned node-attention mechanism.
Key results
- On summer→winter, FID 75.8 vs 81.8 for HnegSRC (best baseline); on apple→orange, FID 139.1 vs 154.1 for NEGCUT.
- On horse→zebra the method matches HnegSRC on FID (34.5 vs 34.4) but achieves better KID (0.271 vs 0.438).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the graph contrastive method instead of HnegSRC | Image translation · Horse→Zebra | KID: 0.438 -> 0.271 | Helped |
| Use the graph contrastive method instead of HnegSRC | Image translation · Label→Cityscape | KID: 0.662 -> 0.605 | Helped |
| Use the graph contrastive method instead of HnegSRC | Image translation · Map→Satellite | FID: 49.2 -> 45.9 | Helped |
| Use the graph contrastive method instead of HnegSRC | Image translation · Summer→Winter | FID: 81.8 -> 75.8 | Helped |
| Use the graph contrastive method instead of HnegSRC | Image translation · Apple→Orange | FID: 158.3 -> 139.1 | Helped |
| Use 1 GNN hop instead of 2 (ablation) | Patch-Wise Graph Contrastive Learning · Horse→Zebra | FID: 34.5 -> 37.9 | Hurt |
| Use 3 GNN hops instead of 2 (ablation) | Patch-Wise Graph Contrastive Learning · Horse→Zebra | FID: 34.5 -> 39.9 | Hurt |
| Set similarity threshold to 0.0 instead of 0.1 (ablation) | Patch-Wise Graph Contrastive Learning · Horse→Zebra | KID: 0.271 -> 0.551 | Hurt |
| Set similarity threshold to 0.4 instead of 0.1 (ablation) | Patch-Wise Graph Contrastive Learning · Horse→Zebra | FID: 34.5 -> 36.8 | Hurt |
| Set similarity threshold to 0.6 instead of 0.1 (ablation) | Patch-Wise Graph Contrastive Learning · Horse→Zebra | FID: 34.5 -> 38.3 | Hurt |
| Add graph pooling (ablation) | Patch-Wise Graph Contrastive Learning · Horse→Zebra | FID: 37.6 -> 34.5 | Helped |
| Use 2 pooling layers instead of 1 (ablation) | Patch-Wise Graph Contrastive Learning · Horse→Zebra | FID: 34.5 -> 35.0 | Hurt |
Limitations
- Gains on horse→zebra and Label→Cityscape are marginal, with FID similar to HnegSRC and not better.
- Ablations show performance degrades with more hops, higher thresholds, or excessive pooling, indicating sensitivity to graph hyperparameters.
Experimental setup
Datasets: horse2zebra Label2Cityscapes map2satellite summer2winter apple2orange
Benchmarks: horse2zebra Label2Cityscapes map2satellite summer2winter apple2orange
Metrics: FID KID
Baselines: CycleGAN MUNIT DistanceGAN GcGAN CUT NEGCUT SeSim Hneg-SRC STROTSS WCT2
Builds on: CUT TAGConv top-K graph pooling LSGAN HnegSRC
Closest papers by meaning
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.87)
- TopoGCL: Topological Graph Contrastive Learning (AAAI 2024 · similarity 0.85)
- Single-View Graph Contrastive Learning with Soft Neighborhood Awareness (AAAI 2025 · similarity 0.84)
- Contrastive Auxiliary Learning with Structure Transformation for Heterogeneous Graphs (AAAI 2025 · similarity 0.84)
- Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing (AAAI 2025 · similarity 0.84)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.84)
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.