Research map / Detection and segmentation / CNN features
HierAmp: Coarse-to-Fine Autoregressive Amplification for Generative Dataset Distillation
Lin Zhao, Xinru Jiang, Xi Xiao, Qihui Fan, Lei Lu, Yanzhi Wang, Xue Lin, Octavia Camps, Pu Zhao, Jianyang Gu
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 class-token attention amplification at all scales (ablation) | VAR distillation baseline · ImageNet-1K IPC=10 | Top-1 accuracy (%): 45.6 -> 47.6 | Improved |
| amplify only coarse scales (5-50%) versus only fine scales (ablation) | HierAmp · ImageNet-1K IPC=10 | Top-1 accuracy (%): Fine 46.5 vs Coarse 47.6 | Improved |
| amplify mid scales only at 5-50% (ablation) | HierAmp · ImageNet-1K IPC=10 | Top-1 accuracy (%): 45.6 -> 46.9 | Improved |
| amplify fine scales only at 7-50% (ablation) | HierAmp · ImageNet-1K IPC=10 | Top-1 accuracy (%): 45.6 -> 46.3 | Improved |
| raise amplification from 3 regions (50%) to 5 regions (50%), full stages (ablation) | HierAmp · ImageNet-1K IPC=10 | Top-1 accuracy (%): 47.1 -> 47.6 | Improved |
| increase top-region share from 30% to 70% (5 regions) at coarse stage (ablation) | HierAmp · ImageNet-1K IPC=10 | Top-1 accuracy (%): 46.7 -> 47.0 | No clear effect |
| use very small coarse-stage amplification (0.1 vs 5) (ablation) | HierAmp · ImageNet-1K IPC=10 | Top-1 accuracy (%): 0.1-5-5: 45.9; 5-0.1-5: 46.6 | Hurt |
| use HierAmp instead of D3HR on cross-arch ImageNet-1K IPC=10 (ResNet-18) | HierAmp · ImageNet-1K IPC=10 cross-arch | Top-1 accuracy (%): 44.3 -> 44.6 | Improved |
| amplification effect on token diversity: coarse scales vs fine scales (ablation) | HierAmp · ImageNet-1K IPC=50 token statistics | codebook token entropy / coverage | Mixed |
Closest papers by meaning
- Data-efficient Large Vision Models through Sequential Autoregression (ICML 2024 · similarity 0.86)
- Synthesize Diagnose and Optimize: Towards Fine-Grained Vision-Language Understanding (CVPR 2024 · similarity 0.86)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.85)
- Flattening the Parent Bias: Hierarchical Semantic Segmentation in the Poincare Ball (CVPR 2024 · similarity 0.85)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.85)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · 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.
