LVFace performance vs. ArcFace/ResNet

About this article

I’m looking at swapping my current face recognition stack for LVFace (the ByteDance paper from ICCV 2025) and wanted to see if anyone has real-world benchmarks yet. Currently, I’m running a standard InsightFace-style pipeline: SCRFD (det_10g) feeding into the Buffalo_L (ArcFace) models. It’s reliable, and I've tuned it to run quickly and with predictable VRAM usage in a long-running environment, but LVFace uses a Vision Transformer (ViT) backbone instead of the usual ResNet/CNN setup, and it ...