Research map / Image generation / Super-resolution

Hardware-Friendly Token-Group Activation Quantization for Low-Bit Mamba Super-Resolution

NeurIPS 2026 · Image generation › Super-resolution · approach: Perceptual and real-world SR

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Token-Group Quantization, a static post-training quantization method for low-bit Mamba super-resolution that predicts token groups from pre-scan features and uses offline group-wise clipping-bound tables, refined with a Fourier amplitude/phase distillation loss (FPR).

Problem

Low-bit post-training quantization of Mamba-based super-resolution models degrades because selective scan produces token-wise activation range heterogeneity that a single layer-wise clipping range cannot fit, while post-scan grouping is hardware-unfriendly for static deployment.

Approach

Approximate a post-scan K-means grouping oracle with a frozen lightweight pre-scan predictor trained on quantization-aware pseudo-labels, select from offline-calibrated group-wise clipping-bound tables, and refine the bounds with a teacher-supervised Fourier amplitude and phase loss.

Key results

• On MambaIRv2-Light x2 W4A4, Ours reaches 33.24 dB on Set14 and 31.34 dB on Urban100, versus 32.45 and 30.34 for layer-wise PTQ.
• Ours+LoRA at W4A4 x4 on BSD100 drops only 0.14 dB from full precision, nearly recovering FP performance.
• Achieves 2.23x speedup over FP16 on Jetson Orin Nano and 1.70x on RTX A6000 for W4A4 static inference.

Surprising findings

• The static pre-scan predicted grouping recovers 82.3% and 81.3% of the gain of the hardware-unfriendly post-scan oracle on Set14 and Urban100.
• The post-scan oracle gives no speedup on either device (0.71x on A6000, 0.98x on Jetson), since it needs an extra observation pass.

Limitations

• Pre-scan prediction and group-wise bound lookup add a small inference overhead over the base quantized path.
• Predictor reaches only about 84% top-1 agreement with pseudo-labels, and gains are smaller on SwinIR where no scan amplification exists.
• Evaluation on other on-device accelerators (NPUs, DSPs, mobile GPUs) is left to future work.

Experimental setup

Datasets: DF2K DIV2K Set5 Set14 BSD100 Urban100 Manga109 ImageNet CBSD68 Kodak24 McMaster Classic5

Benchmarks: Set5 Set14 BSD100 Urban100 Manga109 ImageNet CBSD68 Kodak24 McMaster Classic5 LIVE1

Metrics: PSNR SSIM Top-1 accuracy latency

Baselines: PTQ4VM Quamba MambaQuant OAQ-SR AIQViT SPR^2Q 2DQuant FIMA-Q APHQ-ViT SpinQuant Hadamard Affine

Builds on: MambaIRv2 MambaIR OAQ-SR SPR^2Q PTQ4VM Quamba MambaQuant Atom

Closest papers by meaning

• MambaIRv2: Attentive State Space Restoration (CVPR 2025 · similarity 0.86)
• GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.85)
• MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.84)
• 2DQuant: Low-bit Post-Training Quantization for Image Super-Resolution (NeurIPS 2024 · similarity 0.84)
• GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.83)
• AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · similarity 0.83)