Research map / Architectures and efficiency / Quantization
SSVQ: Unleashing the Potential of Vector Quantization with Sign-Splitting
Shuaiting Li, Juncan Deng, Chengxuan Wang, Kedong Xu, Rongtao Deng, Hong Gu, Haibin Shen, 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 learnable sign bit and freezing, SSVQ instead of VQ | VQ · DeiT-tiny ImageNet (21x) | top-1 accuracy: +12% absolute | Improved |
| add learnable signs instead of fixed sign (ablation) | SSVQ (fixed sign) · DeiT-tiny ImageNet | top-1 accuracy: +3% to 9% | Improved |
| add iterative freezing of learnable signs (ablation) | SSVQ (no freezing) · DeiT-tiny ImageNet | top-1 accuracy: +5% | Improved |
| add learnable signs plus freezing to fixed-sign baseline (ablation) | SSVQ (fixed sign) · DeiT-tiny ImageNet (21x) | top-1 accuracy: +14.8% | Improved |
| use freeze interval 500 instead of 100 (ablation) | SSVQ · DeiT-T (k=16,d=8) | top-1 accuracy: 60.9 -> 61.90 | Improved |
| use freeze interval 500 instead of 1000 (ablation) | SSVQ · DeiT-T (k=16,d=8) | top-1 accuracy: 61.55 -> 61.90 | Improved |
| use majority sign voting instead of EMA criterion (ablation) | SSVQ · DeiT-T (k=16,d=8) | top-1 accuracy: 60.8 -> 61.90 | Improved |
| use SSVQ instead of VQ | VQ · VOC segmentation | mIoU: 65.7 -> 67.7 | Improved |
| use SSVQ instead of VQ | VQ · COCO YOLOv9-S | bbox AP50:95: 30.0 -> 35.2 | Improved |
| use SSVQ instead of VQ | VQ · Stable Diffusion v1-4 COCO prompts | FID-to-FP: 24.5 -> 13.6 | Improved |
| use SSVQ instead of VQ | VQ · Llama3.2-1B 2-bit Wiki2 | perplexity: 63.1 -> 13.9 | Improved |
| train full (codebook+BN+FC) rather than only BN+FC in conventional VQ (ablation) | VQ · MobileNet-V2 ImageNet | top-1 accuracy: 50.01 -> 55.27 | Mixed |
Later papers that compared with this method
1 later accepted paper reports SSVQ as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- VQT: Vector Quantization Tuning for Efficient Fine-tuning and Compression of Pre-trained Vision Transformers (ECCV 2026 · 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
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.88)
- RSAVQ: Riemannian Sensitivity-Aware Vector Quantization for Large Language Models (NeurIPS 2025 · similarity 0.87)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.87)
- Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression (NeurIPS 2025 · similarity 0.87)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.87)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 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.
