yejunliang23/Nano3D-Edit-100k

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- 3d
- 3d-edit
- nano-banana
- text-to-3d
size_categories:
- 100K<n<1M
---

# Nano3D-Edit-100k

This dataset is the official data release for **Nano3D**, a training-free framework for precise and coherent 3D object editing without masks.


> **Paper:** [Nano3D: A Training-Free Approach for Efficient 3D Editing Without Masks](https://arxiv.org/abs/2510.15019)
> **Project Page:** [https://jamesyjl.github.io/Nano3D/](https://jamesyjl.github.io/Nano3D/)

Nano3D integrates FlowEdit into TRELLIS to perform localized 3D edits guided by front-view renderings, and introduces Voxel/Slat-Merge strategies to preserve structural consistency between edited and unedited regions. This dataset was constructed using the Nano3D pipeline (image → 3D → edit), which avoids render-projection artifacts and yields highly consistent editing pairs.

---

## Dataset Structure

```
Nano3D-Edit/
├── v0_15k/ # Early test version (~15k pairs, human-curated)
│ ├── editing_assets/
│ └── info/
└── v1_100k/
├── editing_assets/
│ ├── part_000.tar.gz # uid_0 ~ uid_999
│ ├── part_001.tar.gz # uid_1000 ~ uid_1999
│ └── ...
└── info/
└── v1_100k_info.jsonl
```

### v0_15k

An early test version of the dataset that has undergone manual curation for quality. Note that the editing instructions for this split were unfortunately lost and are not available.

### v1_100k

The official full-scale release containing **100,000 high-quality 3D editing pairs**. Each sample is stored in a folder `uid_{i}` (i = 0 to 99999) under `editing_assets/`:

| File | Description |
|------|-------------|
| `source.png` | Front-view image of the original 3D asset |
| `edit.png` | Front-view image after 2D editing |
| `src_mesh.glb` | Original 3D mesh |
| `tar_mesh.glb` | Edited 3D mesh |
| `src_slat.pt` | Sparse latent (SLAT) of the original asset |
| `tar_slat.pt` | Sparse latent (SLAT) of the edited asset |
| `edit_voxel_post.ply` | Post-processed edit voxel *(not always present)* |

#### info/

The `info/` folder contains `v1_100k_info.jsonl`, a JSONL file with one record per sample. Each line has the following fields:

| Field | Description |
|-------|-------------|
| `uid` | Unique sample identifier, e.g. `uid_0` |
| `prompt` | Detailed text description of the original 3D asset |
| `category` | Object category, e.g. `Plant`, `Vehicle`, `Furniture` |
| `edit_instruction` | Natural language instruction describing the edit to apply |

Example entry:

```json
{
"uid": "uid_0",
"prompt": "A Victorian-style ornamental potted plant featuring stiff, symmetrical fronds and tightly coiled fern-like leaves carved from solid obsidian, its glossy black surface pockmarked with irregular rust-colored corrosion holes that bleed oxidized orange-brown mineral deposits into the surrounding grooves, evoking aged maritime decay reminiscent of abandoned Marine base flora in the One Piece world.",
"category": "Plant",
"edit_instruction": "Add a flared, ribbed gramophone horn and a side-mounted mechanical crank."
}
```

---

## Editing Types

The dataset currently contains two types of edits:

- **add** — adding a new object or part to the asset
- **replace** — replacing an existing part of the asset

> **To obtain `remove` samples:** The `remove` editing type is not directly included. You can construct it from `add` instructions by reversing them with an LLM — e.g., *"add a hat on the head"* → *"remove the hat from the head"*.

---

## Citation

If you use this dataset, please cite:

```bibtex
@article{ye2025nano3d,
title={NANO3D: A Training-Free Approach for Efficient 3D Editing Without Masks},
author={Ye, Junliang and Xie, Shenghao and Zhao, Ruowen and Wang, Zhengyi and Yan, Hongyu and Zu, Wenqiang and Ma, Lei and Zhu, Jun},
journal={arXiv preprint arXiv:2510.15019},
year={2025}
}
```