Research map / Architectures and efficiency / Pruning and efficient tuning
EffiDINO: Task-Specific Model Pruning via Gram Anchoring Subspace Consistency
Jianjian Yin, Liulei Li, Tao Chen, Yi Chen, Yazhou Yao, Wenguan Wang
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 basis-agnostic constraint, residual constraint and task-specific adaptation to the dense-alignment baseline (ablation) | Cut-ViT · ADE20K (20% sparsity) | mIoU: 48.4 -> 49.7 | Improved |
| add spatial and channel gram anchoring to the baseline (ablation) | Cut-ViT · ADE20K (20% sparsity) | mIoU: 48.4 -> 50.0 | Improved |
| use spectral entropy weighting instead of static weighting (ablation) | Cut-ViT · COCO detection (30% sparsity) | mAP: 47.9 -> 48.6 | Improved |
| use spectral entropy weighting instead of static weighting (ablation) | Cut-ViT · DAVIS-2017 VOS (30% sparsity) | (J&F)m / Jm / Fm: 58.8 / 57.2 / 60.3 -> 59.4 / 58.0 / 60.8 | Improved |
| use spectral entropy weighting instead of static weighting (ablation) | Cut-ViT · FG3DCar / JODS / SBD matching (30% sparsity) | PCK@0.05: 70.2 / 55.5 / 37.3 -> 70.8 / 56.3 / 37.9 | Improved |
| replace MSE alignment with basis-invariant subspace consistency (ablation) | Cut-ViT · COCO detection | mAP: 54.7 -> 55.7 | Improved |
| add MSE alignment to the baseline (ablation) | Cut-ViT baseline · COCO detection | mAP: 54.3 -> 54.7 | Improved |
| increase principal components K from 96 to 192 (ablation) | Cut-ViT · ADE20K (20% sparsity) | mIoU: 49.6 -> 50.0 | Improved |
| increase principal components K from 192 to 500 (ablation) | Cut-ViT · ADE20K (20% sparsity) | mIoU: 50.0 -> 50.2 | No clear effect |
| prune on the target dataset instead of ImageNet (ablation) | SNOWS · DAVIS-2017 VOS (30% sparsity) | (J&F)m: 52.9 -> 53.4 | No clear effect |
| prune on the target dataset instead of ImageNet (ablation) | SnapViT · DAVIS-2017 VOS (30% sparsity) | (J&F)m: 56.0 -> 56.4 | No clear effect |
| apply Cut-ViT instead of SnapViT on a SAM backbone | SAM · Promptable segmentation (30% sparsity) | score: 68.6 -> 71.3 | Improved |
Closest papers by meaning
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.88)
- Mostly Text, Smart Visuals: Asymmetric Text-Visual Pruning for Large Vision-Language Models (CVPR 2026 · similarity 0.88)
- When and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLA (NeurIPS 2026 · similarity 0.87)
- IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning (ICLR 2026 · similarity 0.87)
- ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models (CVPR 2025 · similarity 0.87)
- VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm (CVPR 2026 · similarity 0.87)
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.
