Research map / Architectures and efficiency / Quantization
OuroMamba: A Data-Free Quantization Framework for Vision Mamba
Akshat Ramachandran, Mingyu Lee, Huan Xu, Souvik Kundu, Tushar Krishna
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 patch-level contrastive loss L_C in place of patch-similarity loss L_PSE (ablation) | OuroMamba-Gen (loss with L_O) · ImageNet-1K Vim-S W4A8 | top-1 accuracy: 73.45 -> 79.81 | Improved |
| add output loss L_O to contrastive loss (ablation) | OuroMamba-Gen · ImageNet-1K Vim-S W4A8 | top-1 accuracy: 75.52 -> 79.81 | Improved |
| add output loss L_O to patch-similarity loss (ablation) | OuroMamba-Gen · ImageNet-1K Vim-S W4A8 | top-1 accuracy: 71.68 -> 73.45 | Improved |
| add L_PSE on top of L_C + L_O (ablation) | OuroMamba-Gen · ImageNet-1K Vim-S W4A8 | top-1 accuracy: 79.81 -> 74.32 | Hurt |
| use only output loss L_O for generation (ablation) | OuroMamba-Gen · ImageNet-1K Vim-S W4A8 | top-1 accuracy: 21.65 vs 75.52 for L_C only | Hurt |
| calibrate with ViT-generated synthetic data instead of real data | QMamba · ImageNet-1K Vim-S W4A8 | top-1 accuracy: 7-14% accuracy drop | Hurt |
| use delta-based weighting instead of uniform weighting for patched state (ablation) | OuroMamba-Gen · ImageNet-1K VMamba-B W4A4 | top-1 accuracy | Improved |
| vary scanning direction used for generation (ablation) | OuroMamba-Gen · ImageNet-1K VMamba-B W4A4 | top-1 accuracy | No clear effect |
| increase synthetic batch size beyond 128 (ablation) | OuroMamba-Gen · ImageNet-1K Vim-S W4A8 | top-1 accuracy | No clear effect |
| use 5x5 neighborhood with top-12 positive patches (ablation) | OuroMamba-Gen · ImageNet-1K Vim-S W4A8 | top-1 accuracy | Improved |
| set periodic outlier list refresh to 10 time-steps (ablation) | OuroMamba-Quant · W4A4 Vim-B and VMamba-B | normalized speedup | Improved |
| remove outlier list refresh (n_refresh = Full) (ablation) | OuroMamba-Quant · W4A4 Vim-B and VMamba-B | speedup | Hurt |
Closest papers by meaning
- MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods (ICLR 2025 · similarity 0.89)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.87)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.86)
- Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (CVPR 2026 · similarity 0.86)
- Mamba-Reg: Vision Mamba Also Needs Registers (CVPR 2025 · similarity 0.86)
- Breaking Modality Heterogeneity in Low-Bit Quantization for Large Vision-Language Models (NeurIPS 2026 · 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.
