LSGQuant: Layer-Sensitivity Guided Quantization for One-Step Diffusion Real-World Video Super-Resolution
LSGQuant introduces a layer-sensitivity guided quantization method for one-step diffusion-based video super-resolution. It overcomes challenges with low-bit quantization by using a Dynamic Range Adaptive Quantizer, a Variance-Oriented Layer Training Strategy, and Quantization-Aware Optimization. This results in near full-precision performance while significantly reducing model size and computational cost for real-world applications.