THU-MIG/RepViT

Health v2 · maintenance only

Adoption not part of the score

1108 stars · 83 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Official PyTorch implementation of RepViT, a family of lightweight CNNs designed by integrating efficient ViT architectural designs into MobileNetV3, and RepViT-SAM, a real-time Segment Anything variant. It provides pretrained models achieving state-of-the-art accuracy-latency trade-offs on mobile devices, including over 80% ImageNet top-1 accuracy at 1.0 ms latency on iPhone 12.

Use cases

• run segment anything in real time on mobile devices
• lightweight image classification model for edge deployment
• replace SAM image encoder with a faster CNN backbone
• deploy vision models on iPhone with Core ML
• pretrained lightweight backbone for downstream vision tasks
• compare lightweight CNN vs ViT architectures

When to choose

• you need SAM-style segmentation with near real-time latency on resource-constrained devices
• you want a lightweight CNN backbone that outperforms lightweight ViTs on mobile latency
• you need pretrained ImageNet models with excellent accuracy-latency trade-offs

When to avoid

• you need a general-purpose vision framework rather than specific model architectures
• you require transformer-based backbones with self-attention
• your deployment targets are not latency-sensitive or mobile

Facets

library · maturity stable

machine-learning deep-learning computer-vision image-processing computer-vision deep-learning machine-learning image-processing python cross-platform vision-transformer cnn segment-anything mobile-deployment image-classification semantic-segmentation cvpr-2024 pytorch model-architecture repvit gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="THU-MIG/RepViT")