---
license: apache-2.0
tags:
- pose-estimation
- vitpose
- mlx
- mlx-swift
- on-device
- apple-silicon
- keypoint-detection
library_name: mlx
base_model: usyd-community/vitpose-base-simple
---

# ViTPose base-simple — MLX

ViTPose (`vitpose-base-simple`) converted to **MLX** for on-device human pose
estimation on Apple Silicon. Weights are float16.

Built for **[MLXPose](https://github.com/NazarKozak/MLXPose)** — a native MLX Swift
ViTPose implementation. The Swift forward pass is numerically verified against the
Hugging Face reference (heatmaps `max|Δ|=1.5e-6`, decoded keypoints `max 3e-5 px`).

- Backbone: plain ViT-base (12 layers, dim 768), patch 16, input 256×192.
- Head: simple decoder → 17 COCO keypoint heatmaps (64×48).
- Conversion: [`convert_vitpose_to_mlx.py`](https://github.com/NazarKozak/MLXPose/blob/main/scripts/convert_vitpose_to_mlx.py).

## Files
- `weights.safetensors` — MLX float16 weights.
- `config.json` — original ViTPose config.

## License
Apache-2.0. Pretrained weights derive from COCO/MPII training data — review dataset
terms for your use case.