Research map / Vision-language and multimodal / CLIP and zero-shot
Art Beyond Semantics: Sheaf-Informed Contrastive Learning for Multi-Relational Representations
Ludovica Schaerf, Antonio Purificato, Piera Riccio, Fabrizio Silvestri, Noa Garcia
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove relation-conditioned modulation (FiLM restriction maps) (ablation) | CANVAS · HertzianaDP | I2T Recall@10: 0.514 -> 0.002 | Hurt |
| replace restriction maps with simple per-relation linear heads (ablation) | CANVAS · HertzianaDP | I2T Recall@10: 0.514 -> 0.170 | Hurt |
| add KL graph-regularized loss to InfoNCE (ablation) | CANVAS · HertzianaDP | T2I Recall@10: 0.004 -> 0.158 | Improved |
| add InfoNCE loss to KL loss (ablation) | CANVAS · HertzianaDP | I2T Recall@10: 0.490 -> 0.514 (full); InfoNCE alone 0.541 | Mixed |
| CANVAS vs best baseline in image-to-text retrieval | CANVAS · HertzianaDP, SemArt+, WikiArt+ | I2T Recall@10: 0.084 -> 0.514 (HertzianaDP); 0.150 -> 0.714 (SemArt+); 0.209 -> 0.781 (WikiArt+) | Improved |
| CANVAS vs best baseline in text-to-image retrieval | CANVAS · WikiArt+ | T2I Recall@10: 0.217 vs 0.307 (SigLIP-ft) | Hurt |
| add relation type to CLIP and SigLIP text prompts | CLIP, SigLIP · WikiArt+ T2I | Recall@5: 0.131 -> 0.100 (CLIP); 0.239 -> 0.169 (SigLIP) | Hurt |
| use predicted relation type instead of ground-truth at inference (ablation) | CANVAS · HertzianaDP, SemArt+, WikiArt+ | Recall: HertzianaDP I2T R@10 0.514 -> 0.578; SemArt+ T2I R@10 0.489 -> 0.414 | No clear effect |
| CANVAS vs baselines on relation-aware classification | CANVAS · SemArt+ and WikiArt+ Style | accuracy: 0.494 -> 0.637 (SemArt+ Style); 0.702 -> 0.790 (WikiArt+ Style) | Improved |
| CANVAS vs ArtSAGENet on HertzianaDP artist classification | CANVAS · HertzianaDP Artist | accuracy: 0.808 vs 0.472 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ColPali — ColPali: Efficient Document Retrieval with Vision Language Models (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PoSh: Using Scene Graphs to Guide LLMs-as-a-Judge for Detailed Image Descriptions (ICLR 2026 · similarity 0.85)
- Scene Graph Thinking: Reinforcing Structured Visual Reasoning for Multimodal Large Language Models (ICML 2026 · similarity 0.85)
- Object-centric binding in Contrastive Language-Image Pretraining (NeurIPS 2025 · similarity 0.85)
- Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.85)
- The Photographer's Eye: Teaching Multimodal Large Language Models to See, and Critique Like Photographers (CVPR 2025 · similarity 0.84)
- Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms (NeurIPS 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.
