Research map / Architectures and efficiency / Pruning and efficient tuning
VQT: Vector Quantization Tuning for Efficient Fine-tuning and Compression of Pre-trained Vision Transformers
Yinglong Li, jiyuan Xia, Jingcheng Xie, Zhiwei Xiong
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 LoRA with STE and codeword reassignment to codebook-only VQ (ablation) | VQ fine-tuning · VTAB-1K ViT-B 1-bit | accuracy (%): 55.1 -> 57.6 | Improved |
| add LoRA with STE and codeword reassignment to codebook-only VQ (ablation) | VQ fine-tuning · VTAB-1K ViT-B 1.5-bit | accuracy (%): 64.4 -> 69.0 | Improved |
| add stochastic sub-vector replacement to LoRA & STE (ablation) | VQT · VTAB-1K ViT-B 1-bit | accuracy (%): 57.6 -> 63.9 | Improved |
| add stochastic sub-vector replacement to LoRA & STE (ablation) | VQT · VTAB-1K ViT-B 2-bit | accuracy (%): 72.5 -> 72.3 | No clear effect |
| add progressive alignment on top of stochastic replacement (ablation) | VQT · VTAB-1K ViT-B 1-bit | accuracy (%): 63.9 -> 67.5 | Improved |
| add progressive alignment on top of stochastic replacement (ablation) | VQT · VTAB-1K ViT-B 1.5-bit | accuracy (%): 69.4 -> 70.8 | Improved |
| use both continuous-anchored and VQ-anchored modes instead of one (ablation) | VQT · VTAB-1K ViT-B 1-bit | accuracy (%): 66.2 -> 67.5 | Improved |
| replace LoRA with DoRA inside VQT instead of direct VQ+DoRA | VQT · VTAB-1K ViT-B 1-bit with DoRA | accuracy (%): 57.4 -> 67.9 | Improved |
| unified fine-tune-and-quantize VQT instead of post-training VQ4DiT on fine-tuned LoRA | VQT · VTAB-1K ViT-B 1-bit | accuracy (%): 60.9 -> 67.5 | Improved |
| unified VQT instead of VQ4DiT with task loss at 2-bit | VQT · VTAB-1K ViT-B 2-bit | accuracy (%): 72.9 -> 72.6 | No clear effect |
| allow codeword reassignment instead of fixed assignments (PQF) | VQT · VTAB-1K ViT-B 1.5-bit | accuracy (%): 65.7 -> 70.8 | Improved |
| use VQT instead of SSVQ on ViT-L at 1 bit | VQT · VTAB-1K ViT-L | accuracy (%): 59.2 -> 66.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SSVQ — SSVQ: Unleashing the Potential of Vector Quantization with Sign-Splitting (ICCV 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.90)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.90)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.90)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.90)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.89)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.89)
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.
