Research map / Architectures and efficiency / Quantization
Efficient Quantization-Aware Adaptation for Visual Foundation Models
Yinglong Li, Xiaoyu Liu, Yutong Liu, Yueyi Zhang, 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 |
|---|---|---|---|
| split a sub-network side adapter and exclude the frozen backbone from the computation graph (ablation) | EQuA on SAM-H W4A4 · SAM-H training memory | training memory (GB): 67.6 -> 18.6 | Improved |
| add gradient recomputation through the frozen backbone (ablation) | EQuA SSA · Kvasir (SAM-H, W4A4) | S-measure: 0.881 -> 0.910 | Improved |
| add LoRA to SSA linear layers (ablation) | EQuA · Kvasir (SAM-H, W4A4) | S-measure: 0.910 -> 0.906; params 29.90M -> 7.67M | No clear effect |
| add block-wise activation quantizer fine-tuning (BAQF) (ablation) | EQuA · Kvasir (SAM-H, W4A4) | S-measure: 0.906 -> 0.915 | Improved |
| perform BAQF (K=10 vs none) (ablation) | EQuA · Leaf (SAM-H, W4A4) | mIoU: 0.699 -> 0.709 | Improved |
| increase SSA dimension 32 to 64 (ablation) | EQuA · Kvasir (SAM-H, W4A4) | E-measure: 0.933 -> 0.947 | Improved |
| increase SSA dimension 64 to 128 (ablation) | EQuA · Kvasir (SAM-H, W4A4) | E-measure: 0.947 -> 0.943 | No clear effect |
| increase LoRA rank 4 to 8 (ablation) | EQuA · Kvasir (SAM-H, W4A4) | S-measure: 0.915 -> 0.919 | No clear effect |
| increase BAQF iterations 100 to 200 (ablation) | EQuA · Leaf (SAM-H, W4A4) | mIoU: 0.709 -> 0.710 | No clear effect |
| EQuA vs QST under W4A4 (merged adapter, quantizer tuning) | SAM-H · Leaf (SAM-H, W4A4) | mIoU: 0.656 -> 0.709 | Improved |
| EQuA vs PEQA on memory, with similar accuracy | SAM-H W4A4 · SAM-H training memory | training memory (GB): 69.4 -> 18.5 | Improved |
| EQuA vs PEQA on accuracy | ViT-B W4A4 · VTAB-1K Group Average | top-1 accuracy: 72.9 -> 72.0 | Hurt |
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.
- NIPQ — NIPQ: Noise Proxy-Based Integrated Pseudo-Quantization (CVPR 2023) · better on 1 of 1 shared result
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
- GANQ: GPU-Adaptive Non-Uniform Quantization for Large Language Models (ICML 2025 · similarity 0.88)
- Q-VLM: Post-training Quantization for Large Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- Time- Memory- and Parameter-Efficient Visual Adaptation (CVPR 2024 · similarity 0.87)
- Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (CVPR 2026 · similarity 0.87)
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.87)
- KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization (NeurIPS 2024 · 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.
