Mamba-based super-resolution models have emerged as efficient substitutes for Transformer-based restoration backbones. However, applying low-bit post-training quantization to them is not straightforward. The difficulty stems from the selective scan mechanism, Mamba's input-dependent state-space operation, which does not behave like standard attention layers during quantization.
To address this, the paper introduces a hardware-friendly token-group activation quantization approach. Rather than quantizing activations uniformly, the method groups tokens in a way that aligns with the selective scan's structure, making the quantization process more tractable on real hardware.
The work is positioned as a post-training solution, meaning it avoids the expense of full retraining. The abstract does not report specific numerical results, so the practical gains remain to be detailed in the full paper. Still, the proposal directly targets a known bottleneck for deploying efficient Mamba-based restoration models on resource-constrained devices.