ECCV 2026
Authors: Kartik Narayan, Vishal M. Patel
Affiliation: Johns Hopkins University

Low-resolution face recognition remains challenging due to severe degradations in probe images, domain differences between high-resolution gallery and low-resolution probe data, and catastrophic forgetting during low-resolution adaptation. FaceMoE introduces a transformer with Mixture-of-Experts feed-forward blocks and a top-k router that dynamically activates specialized experts for different semantic facial regions. This resolution-aware sparse routing improves feature extraction under degradation while preserving pretrained knowledge and scaling capacity efficiently.

Across eleven datasets (high-quality, mixed-quality, and low-resolution benchmarks), FaceMoE outperforms prior state-of-the-art methods, including strong gains on BRIAR Protocol 3.1, IJB-S, and TinyFace.

The motivation figure highlights three core LR-FR challenges: (1) degraded probe frames contain weak identity cues, making feature aggregation difficult; (2) a strong HR gallery vs LR probe domain gap changes which facial regions are discriminative; and (3) naive low-resolution fine-tuning can cause catastrophic forgetting.

FaceMoE addresses these challenges by introducing sparse expert FFNs, routing each token through top-k specialized experts, and improving adaptation to low-resolution data with minimal drop on high-quality and mixed-quality benchmarks.

FaceMoE replaces the standard transformer FFN with multiple expert MLPs and a learnable top-k router. Tokens are sparsely routed to expert subsets, enabling resolution-aware feature extraction from different semantic facial regions. A composite objective with CosFace loss, router z-loss, and load-balancing loss stabilizes expert specialization. The reported effective configuration is N = 3 experts with k = 2 active experts per token.

You can download the weights using:

Coming soon ...

Check our GitHub repo for complete training and inference instructions: https://github.com/Kartik-3004/FaceMoE