Fast-Converging and Architecture-Agnostic 6-Bit Face Recognition via Gradient Coordination and Refined Data

Abstract

Deploying face recognition models on resource-constrained Edge AI devices requires aggressive model compression, where low-bit quantization is a promising solution. However, extreme 6-bit (Q6) quantization often distorts the embedding space, severely degrading recognition performance. In this paper, we propose an architecture-agnostic gradient-coordinated quantization framework that stabilizes training by aligning the student network’s optimization trajectory with a full-precision teacher. Specifically, we introduce a β-weighted Mean Squared Error (β-MSE) alignment objective to amplify guidance signals, combined with a three-phase training strategy to stabilize dynamic ranges and optimization. Departing from the conventional reliance on large-scale synthetic datasets, we adopt a refined dataset strategy using approximately 53,500 real-world images enriched with high-quality samples covering diverse pose and age variations. Experimental results on standard benchmarks including LFW, CFP-FP, AgeDB-30, CALFW and CPLFW show that our method achieves 57.4x faster convergence (3,135 vs. 180,000 iterations) compared to real-data-based state-of-the-art approaches like QuantFace when normalized to the same batch size. Our framework recovers nearly all performance loss, achieving a 102% recovery rate on AgeDB-30, and demonstrates robustness across multiple architectures, including iResNet-50 and MobileFaceNet.