Reproduce “QuantFace: Efficient Quantization for Face Restoration” on OpenResearch

About this paper

Jiatong Li, Libo Zhu, Haotong Qin, Jingkai Wang, Linghe Kong, Guihai Chen, Yulun Zhang, Xiaokang Yang

QuantFace is a low-bit quantization framework for one-step diffusion face restoration (OSDFR) models that reduces model parameters by 85% and operations by 83% while preserving high perceptual quality. This framework enables efficient, high-quality face restoration even at 4-bit precision, making advanced models deployable on resource-constrained devices.

Key insights

• Combining rotation and scaling effectively manages extreme outliers and balances inter-channel activation distributions, which is crucial for low-bit quantization in generative models.
• A dual-branch low-rank adaptation approach, integrated with knowledge distillation, provides enhanced optimization flexibility for quantized models to maintain both structural fidelity and fine detail generation.
• Layer-wise sensitivity to quantization can be leveraged through perceptually-aware adaptive bit-width allocation to maximize overall quality within a resource budget, especially in critical layers like downsampling and upsampling.

Summary by alphaXiv.