README.md

---
license: mit
task_categories:
- text-to-3d
- image-to-3d
tags:
- 3d-editing
- shape-editing
- shape-generation
- evaluation-benchmark
- siggraph2026
size_categories:
- n<1K
---

# Prox-E ShapeTalk Evaluation Benchmark Dataset

This is the official benchmark dataset for **Prox-E: Fine-Grained 3D Shape Editing via Primitive-Based Abstractions (SIGGRAPH'26)**.

## 📄 Dataset Overview
This benchmark consists of a subset of random **600 samples** from the **ShapeTalk** dataset. It is used to evaluate identity preservation, 3D quality, and edit fidelity.

## 📂 Repository Structure
The repository is structured to seamlessly plug into the `Prox-E` unified evaluator:

- **`input_shapes/`**: Directory containing `600` input 3D meshes in OBJ format (`.obj`).
- **`rendered_images/`**: 600 high-resolution upright 2D renders (PNG) of the input shapes, produced from TRELLIS's inversion outputs.
- **`point_cloud/`**: Directory containing `600` dense point cloud of the input meshes (`.npz`).
- **`instructions.json`**: Key-value metadata containing evaluation prompts, object classes and part-specific keywords.
- **`shapetalk_600_set.csv`**: Full detailed metadata from the original ShapeTalk dataset on this 600-sample subset.

---

## 🛠️ Usage with Prox-E Evaluation Tool

### 1. Download the Dataset
You can clone this dataset repository using Git LFS:
```bash
git lfs install
git clone https://huggingface.co/datasets/haopt/prox-e-shapetalk-benchmark
```

### 2. Run the Unified Evaluator
Once downloaded, plug this benchmark directly into the Prox-E unified evaluation pipeline to compute metric scores (e.g., l-GD, LPIPS, DINO-I, FID, PFD, CLIP, and VQA):

```bash
python -m evals.main \
  --pred_dir <flat folder of pred .glb/.obj/.ply> \
  --input_dir prox-e-shapetalk-benchmark/input_shapes \
  --instructions_json prox-e-shapetalk-benchmark/instructions.json \
  --input_render_dir prox-e-shapetalk-benchmark/rendered_images \
  --input_pcd_dir prox-e-shapetalk-benchmark/point_cloud \
  --output_dir <eval_run_dir> \
  --device cuda:0 \
  --metrics identity quality fidelity \
  --enable_vqa
```

Refer to the official [Prox-E repo](https://github.com/etaisella/Prox-E/tree/main/evals) for setup and full usage options!

## 📜 Citation
If you use this benchmark dataset in your work, please cite the Prox-E paper and the ShapeTalk dataset:

```bibtex
@inproceedings{sella2026proxefinegrained3dshape,
    title={Prox-E: Fine-Grained 3D Shape Editing via Primitive-Based Abstractions},
    author={Etai Sella and Hao Phung and Nitay Amiel and Or Litany and Or Patashnik and Hadar Averbuch-Elor},
    booktitle={Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers},
    year={2026},
}

@inproceedings{achlioptas2023shapetalk,
    title={{ShapeTalk}: A Language Dataset and Framework for 3D Shape Edits and Deformations},
    author={Achlioptas, Panos and Huang, Ian and Sung, Minhyuk and Tulyakov, Sergey and Guibas, Leonidas},    
    booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},    
    year={2023}
}
```