Research map / Architectures and efficiency / Attention and transformers
When Token Compression Breaks: Structural Pruning vs. Token Reduction for Robust ViT Segmentation under High Compression
Tien-Phat Nguyen, Ngai-Man Cheung
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 |
|---|---|---|---|
| mild token compression with CTS | DeiT-B Segmenter · ADE20K val | mIoU clean/noise: 47.22/35.58 -> 46.03/34.76 | No clear effect |
| mild token compression with CTS | DeiT-B Segmenter · Cityscapes val | mIoU clean/noise: 78.25/63.26 -> 78.15/63.07 | No clear effect |
| aggressive ToMe merging (3.7x FLOPs reduction) | DeiT-B Segmenter · ADE20K val | mIoU clean/noise: 47.22/35.58 -> 23.94/16.86 | Hurt |
| aggressive ToMe merging | DeiT-B Segmenter · Cityscapes val | mIoU clean/noise: 78.25/63.26 -> 64.50/50.56 | Hurt |
| aggressive segmentation-aware token compression (ALGM) | DeiT-B Segmenter · ADE20K val | mIoU clean: 47.22 -> 42.85 | Hurt |
| structural pruning (NViT-S) at 3.8x FLOPs reduction | DeiT-B Segmenter · ADE20K val | mIoU clean: 47.22 -> 45.30 | Hurt |
| mild structural pruning (NViT) at 1.4x | DeiT-B Segmenter · ADE20K val | mIoU clean: 47.22 -> 47.93 | Improved |
| prune-then-merge instead of ToMe alone at high compression (ablation) | NViT-B+ToMe · ADE20K val | mIoU clean: 23.94 -> 45.71 | Improved |
| prune-then-merge instead of NViT-S alone at high compression (ablation) | NViT-B+ToMe · ADE20K val | mIoU noise: 32.86 -> 34.33 | Improved |
| prune-then-merge instead of NViT-S alone (ablation) | NViT-B+ToMe · Cityscapes val | mIoU clean: 76.76 -> 77.52 | Improved |
| prune-then-merge instead of CTS at high compression (ablation) | NViT-B+ToMe · Cityscapes-C | mIoU noise: 60.99 -> 60.71 | Hurt |
| prune-then-merge vs very aggressive pruning at matched FPS (ablation) | NViT-B+ToMe · ADE20K val FPS-matched | mIoU clean: 41.01 -> 45.17 | Improved |
Closest papers by meaning
- Token Cropr: Faster ViTs for Quite a Few Tasks (CVPR 2025 · similarity 0.89)
- ImagePiece: Content-aware Re-tokenization for Efficient Image Recognition (AAAI 2025 · similarity 0.87)
- Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity (ICLR 2026 · similarity 0.86)
- Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers (NeurIPS 2024 · similarity 0.86)
- MADTP: Multimodal Alignment-Guided Dynamic Token Pruning for Accelerating Vision-Language Transformer (CVPR 2024 · similarity 0.86)
- TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models (NeurIPS 2026 · similarity 0.86)
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.
