Research map / Image generation / Autoregressive and latent generation
Learning AND–OR Templates for Compositional Representation in Art and Design
Liaoruxing Zhang, Xin Jin, Wenbo Yuan, ChenYu Fan, Song-Chun Zhu
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 |
|---|---|---|---|
| add late fusion of template configuration features to VGG19 | VGG19 · AVA accuracy | accuracy (%): 78.08 -> 81.43 | Improved |
| add late fusion of template configuration features to ResNet50 | ResNet50 · AVA accuracy | accuracy (%): 85.71 -> 85.87 | Improved |
| add late fusion of template configuration features to ViT-B/16 | ViT-B/16 · AVA accuracy | accuracy (%): 85.38 -> 86.20 | Improved |
| add late fusion of template features to VGG19 | VGG19 · AVA | SRCC: 0.7506 -> 0.7893 | Improved |
| add late fusion of template features to ResNet50 | ResNet50 · AVA | SRCC: 0.7842 -> 0.8156 | Improved |
| add late fusion of template features to ViT-B/16 | ViT-B/16 · AVA | SRCC: 0.7590 -> 0.7855 | Improved |
| use template consistency score instead of a deep 2-class classifier (ResNet50) for correlation with human ratings | AND-OR templates · AVA | SRCC: 0.7842 -> 0.8419 | Improved |
| use templates instead of ResNet50 for AVA accuracy | AND-OR templates · AVA | accuracy (%): 85.71 -> 85.65 | No clear effect |
| use templates instead of CNN baselines on a 10-class Places365 subset | AND-OR templates · Places365 (10-class subset) | Top-5 error (%): 8.61 -> 8.03 | Improved |
| perturb templates by part substitution, relation shuffling or geometric jitter (ablation) | consistency score · CITD themes | consistency score | Hurt |
| increase training images in semi-supervised learning (ablation) | template AND-OR graph · AVA/AADB semi-supervised | number of nodes | Mixed |
| increase training sample size n to 100 in semi-supervised learning (ablation) | semi-supervised templates · cross-validation KL | KL divergence: reduced to 0.1 | Improved |
Closest papers by meaning
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.87)
- Detect Anything in Graphic Design: Element-Level Rewards for Autoregressive Detection (NeurIPS 2026 · similarity 0.86)
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · similarity 0.85)
- Unlocking Compositional Generalization in Continual Few-Shot Learning (NeurIPS 2026 · similarity 0.85)
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 2025 · similarity 0.85)
- Shape of Thought: Progressive Object Assembly via Visual Chain-of-Thought (ICML 2026 · similarity 0.85)
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.
