In this work, we present a dual-stage low-bit post-training quantization (PTQ) method for image super-resolution, namely 2DQuant, which achieves efficient and accurate SR under low-bit quantization.

Different distribution shifts require different algorithmic and operational interventions. Methodological research must be grounded by the specific shifts they address.

We demonstrate the versatility of our approach for both fully data-driven and for physics-aware neural solvers.