Research map / Image generation / Super-resolution
FaceMoE: Mixture of Experts for Low-Resolution Face Recognition
Kartik Narayan, Vishal Patel
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace a single FFN with 3 FFN experts and a top-2 router plus auxiliary losses (ablation) | Swin-B face encoder · TinyFace | Rank-1: 75.10 -> 76.18 | Improved |
| add top-k router and MoE without auxiliary losses (ablation) | FaceMoE · TinyFace | Rank-1: 75.40 -> 74.94 | Hurt |
| add router z-loss and load balancing loss (ablation) | FaceMoE · TinyFace | Rank-1: 74.94 -> 76.18 | Improved |
| use learned routing instead of random expert assignment (ablation) | FaceMoE · TinyFace | Rank-1: 75.40 -> 76.18 | Improved |
| use full-capacity FFN experts instead of LoRA FFN experts (ablation) | FaceMoE · TinyFace | Rank-1: 43.61 -> 76.18 | Improved |
| use multiple experts instead of a single shared MLP expert (ablation) | FaceMoE · TinyFace | Rank-1: 62.71 -> 76.18 | Improved |
| increase number of experts from 3 to 8 (ablation) | FaceMoE · TinyFace | Rank-1: 76.18 -> 2.31 | Hurt |
| use N=3, k=2 instead of N=1 or N=4 (ablation) | FaceMoE · BRIAR Protocol 3.1 | Rank-1: 70.2% (N=1) -> 73.1% (N=3, k=2) | Improved |
| apply FaceMoE to a ViT-B backbone (ablation) | ViT-B · BRIAR Protocol 3.1 | Rank-1: 55.59 -> 71.34 | Improved |
| apply FaceMoE to a ViT-B backbone (ablation) | ViT-B · IJB-B TAR@FAR=1e-4 | TAR: 95.18 -> 89.75 | Hurt |
| scale pretraining data from WebFace4M to WebFace12M (ablation) | FaceMoE · BRIAR Protocol 3.1 | Rank-1: 73.06 -> 74.77 | Improved |
| fine-tune CosFace on BRIAR instead of using the pretrained model | CosFace · BRIAR Protocol 3.1 | TAR@FAR=0.01%: 34.29 -> 11.62 (ViT-B pretrained vs finetuned) | Hurt |
Closest papers by meaning
- LVFace: Progressive Cluster Optimization for Large Vision Models in Face Recognition (ICCV 2025 · similarity 0.87)
- Teaching Where to Look: Attention Similarity Knowledge Distillation for Low Resolution Face Recognition (ECCV 2022 · similarity 0.86)
- ProxyFusion: Face Feature Aggregation Through Sparse Experts (NeurIPS 2024 · similarity 0.84)
- LAFS: Landmark-based Facial Self-supervised Learning for Face Recognition (CVPR 2024 · similarity 0.84)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.84)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
