Research map / Architectures and efficiency / Quantization
ViM-VQ: Efficient Post-Training Vector Quantization for Visual Mamba
Juncan Deng, Shuaiting Li, Zeyu Wang, Kedong Xu, Hong Gu, Kejie Huang
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 fast convex combination optimization to basic VQ (ablation) | ViM-VQ on Vim-T 2-bit · ImageNet-1K (2-bit weights) | top-1 accuracy: 33.18 -> 71.42 | Improved |
| add incremental vector quantization on top of combination optimization (ablation) | ViM-VQ on Vim-T 2-bit · ImageNet-1K (2-bit weights) | top-1 accuracy: 71.42 -> 72.17 | Improved |
| increase candidate codewords n from 4 to 16 (ablation) | ViM-VQ on Vim-T 2-bit · ImageNet-1K (2-bit weights) | top-1 accuracy: 72.17 -> 72.31 | Improved |
| increase candidate codewords n from 16 to 64 (ablation) | ViM-VQ on Vim-T 2-bit · ImageNet-1K (2-bit weights) | top-1 accuracy: 72.31 -> 72.34 | No clear effect |
| use incremental quantization so inference matches calibration accuracy | ViM-VQ vs VQ4DiT on Vim-T 2-bit · ImageNet-1K (2-bit weights) | calibration-to-inference gap: VQ4DiT 71.33 -> 71.04; ViM-VQ 72.17 -> 72.17 | Improved |
| use limited-candidate local convex combination instead of global soft assignment (DKM) | ViM-VQ vs DKM · Vim-T 2-bit calibration | GPU memory: 25 GB -> 11 GB | Improved |
| use customized CUDA kernel for matrix-vector multiplication | ViM-VQ on Vim-B 2-bit · ImageNet-1K val | inference time: 357s -> 328s | Improved |
| apply ViM-VQ instead of DKM at 1-bit | ViM-VQ on VMamba-T · ADE20K (1-bit, single-scale) | mIoU: 37.2 -> 41.1 | Improved |
| apply ViM-VQ instead of DKM at 1-bit | ViM-VQ on VMamba-T · MSCOCO (1-bit) | box AP: 37.5 -> 40.9 | Improved |
| apply ViM-VQ instead of DKM at 3-bit | ViM-VQ on Vim-T · ImageNet-1K (3-bit weights) | top-1 accuracy: 73.64 -> 74.79 | Improved |
| use vector quantization instead of uniform quantization at 2-bit | ViM-VQ family on Vim-T · ImageNet-1K (2-bit weights) | top-1 accuracy: 0.16 -> 33.18 | 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.
- VQ4DiT — VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025) · better on 3 of 3 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
- VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models (CVPR 2026 · similarity 0.89)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.88)
- Stochastic Layer-Wise Shuffle for Improving Vision Mamba Training (ICML 2025 · similarity 0.88)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.87)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.87)
- MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods (ICLR 2025 · 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.
