---
title: EdgeCrafter
emoji: 🪶
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 6.24.0
app_file: app.py
short_description: Detection, segmentation & pose with compact edge ViTs
python_version: "3.12"
startup_duration_timeout: 30m
pinned: false
---

# EdgeCrafter

Demo of **EdgeCrafter: Compact ViTs for Edge Dense Prediction via Task-Specialized
Distillation** — one distilled ViT backbone family serving three dense-prediction
tasks:

| task | models | classes |
|---|---|---|
| Object detection | `Intellindust/ECDet_{S,M,L,X}` | COCO 80 |
| Instance segmentation | `Intellindust/ECSeg_{S,M,L,X}` | COCO 80 |
| Human pose estimation | `Intellindust/ECPose_{S,M,L,X}` | COCO 17 keypoints |

All twelve checkpoints (~10M → ~50M params) are loaded and selectable in the UI.
Preprocessing (640×640 resize + ImageNet normalization), the deploy-mode
re-parameterization, postprocessing and the drawing style are ported 1:1 from the
authors' `tools/inference/torch_inf.py` scripts and `hf_models.ipynb`.

- Paper: https://huggingface.co/papers/2603.18739
- Code: https://github.com/Intellindust-AI-Lab/EdgeCrafter (Apache-2.0)

The `ecdetseg/` and `ecpose/` folders vendor the inference-only subset of the
authors' Apache-2.0 model code.

Example photos are COCO images redistributed from the Apache-2.0 repositories
[DEIMv2](https://github.com/Intellindust-AI-Lab/DEIMv2) (`example.jpg`) and
[DETRPose](https://github.com/SebastianJanampa/DETRPose) (`examples/example1.jpg`,
`examples/example2.jpg`).