Research map / Architectures and efficiency / Attention and transformers
Token Compensator: Altering Inference Cost of Vision Transformer without Re-Tuning
Shibo Jie, Yehui Tang, Jianyuan Guo, Zhi-Hong Deng, Kai Han, Yunhe 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 |
|---|---|---|---|
| plug ToCom into models trained without compression, target r=16 | ToMe on DeiT-B · VTAB-1k Specialized | avg accuracy: 80.0 -> 83.6 | Improved |
| plug ToCom into models trained without compression, target r=16 | ToMe on DeiT-B · VTAB-1k Natural | avg accuracy: 73.7 -> 75.8 | Improved |
| plug ToCom into models trained without compression, target r=16 | ToMe on DeiT-B · VTAB-1k Structured | avg accuracy: 56.7 -> 57.9 | Improved |
| add ToCom with source r=0 on CIFAR100 and FGVC, target r=8 and 16 | ToMe on DeiT-B · CIFAR100 and FGVC average | accuracy: +0.5 (r=8), +1.7 (r=16) | Improved |
| add ToCom for training-time acceleration (target r lower than source r) | ToMe on DeiT-B · CIFAR100, FGVC, VTAB-1k | accuracy | Mixed |
| add ToCom to ToMe for semantic segmentation at target r=16 | ToMe on DeiT-B384 + Segmenter · ADE20k val | mIoU (SS): 41.3 -> 43.4 | Improved |
| use a single shared LoRA (h=32) instead of 16 groups of adjacent-degree LoRA (ablation) | ToCom · VTAB-1k | avg accuracy | Hurt |
| replace self-distillation loss with classification cross-entropy loss (ablation) | ToCom · VTAB-1k (source r > target r) | avg accuracy | Hurt |
| use separate LoRA for inverse direction instead of subtracting (invertibility assumption) (ablation) | ToCom · VTAB-1k | avg accuracy | Hurt |
| increase LoRA hidden dimension h beyond 2 (ablation) | ToCom · VTAB-1k | avg accuracy | No clear effect |
| apply ToCom to EViT token pruning instead of ToMe | EViT · VTAB-1k | avg accuracy | Improved |
Closest papers by meaning
- Inference Optimal VLMs Need Fewer Visual Tokens and More Parameters (ICLR 2025 · similarity 0.88)
- HiRED: Attention-Guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models (AAAI 2025 · similarity 0.87)
- DeepCompress-ViT: Rethinking Model Compression to Enhance Efficiency of Vision Transformers at the Edge (CVPR 2025 · similarity 0.86)
- DUET-VLM: Dual stage Unified Efficient Token reduction for VLM Training and Inference (CVPR 2026 · similarity 0.86)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.86)
- Vision-centric Token Compression in Large Language Model (NeurIPS 2025 · 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.
