---
license: cc-by-4.0
task_categories:
- image-to-image
tags:
- Generative Modeling
- Image Editing
- Geometric Editing
- 3D Vision
size_categories:
  - 100K<n<1M
language:
- en
configs:
  - config_name: 100K-Syn
    data_files:
      - split: train
        path: data/100K-Syn/*.tar
  - config_name: 10K-Objectron
    data_files:
      - split: train
        path: data/10K-Objectron/*.tar
  - config_name: 10K-Syn
    data_files:
      - split: train
        path: data/10K-Syn/*.tar
---

# [NeurIPS-2026] Thinking in Boxes: 3D Editing in Real Images Made Easy

<div>
  <img src="assets/thumbnail.png">
</div>

[**Pradhaan S Bhat**](https://pradhaansbhat.github.io/)<sup>1</sup><sup>∗</sup> · [**Naveen Chandra R**](https://www.linkedin.com/in/naveen-chandra-r-7230aa192)<sup>1</sup><sup>∗</sup> · [**Rishubh Parihar**](https://rishubhpar.github.io/)<sup>1</sup> · [**Vaibhav Vavilala**](https://www.linkedin.com/in/vaibhav-vavilala)<sup>2</sup> · [**R. Venkatesh Babu**](https://cds.iisc.ac.in/faculty/venky/)<sup>1</sup> · [**D.A. Forsyth**](http://luthuli.cs.uiuc.edu/~daf/) · [**Anand Bhattad**](https://anandbhattad.github.io/)<sup>4</sup>

<sup>1</sup> Indian Institute of Science
<sup>2</sup> Apple
<sup>3</sup> UIUC
<sup>4</sup> Johns Hopkins University

<sup>∗</sup> Equal Contribution

[![arXiv](https://img.shields.io/badge/arXiv-2606.20556-b31b1b.svg)](https://arxiv.org/abs/2606.20556)
[![Project Page](https://img.shields.io/badge/Project-Page-1f6feb.svg)](https://thinking-in-boxes.github.io)
[![GitHub](https://img.shields.io/badge/Github-Repository-blue?logo=github)](https://github.com/PradhaanSBhat/Thinking-In-Boxes)
[![Hugging Face Model](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-yellow)](https://huggingface.co/pradhaansbhat/Thinking-In-Boxes)
[![Demo](https://img.shields.io/badge/🎬%20Demo-blue.svg)]()

<div>
  <img src="assets/pipeline.png">
</div>

This is the training dataset used in the paper **Thinking In Boxes: 3D Editing in Real Images Made Easy**

Thinking-In-Boxes is an Image-to-Image Generative model for Geometric Image Editing. 
This dataset consists of images of 1-2 object scenes placed on a floor and rendered from two viewpoints. 
In addition, it includes our scene representation where objects as represented as 3D Coloured Boxes placed on a shaded floor, which disambiguates object and camera transformations.

## Dataset Structure

This dataset contains three independent subsets, each stored as WebDataset `.tar` shards:

| Subset | Folder | Approx. size | Description |
|---|---|---|---|
| 100K-Syn | `data/100K-Syn/` | 100,000 scenes | Synthetic 2-object scenes used in Stage-1 finetuning |
| 10K-Objectron | `data/10K-Objectron/` | 10,000 scenes | Real-world Objectron-derived scenes used in Stage-2 finetuning |
| 10K-Syn | `data/10K-Syn/` | 10,000 scenes | Synthetic 2-object scenes used in Stage-2 finetuning |

Each "scene" (sample) contains 4 files:

- `bbox_0.png`, `bbox_1.png` — scene representations for source and target configurations.
- `rgb_0.png`, `rgb_1.png` — RGB renders representing source and target configurations.

**Note**: The *.png files for the 100K-Syn and 10K-Syn subsets are in 512x512 resolution, and 1440x1920 resolution for the 10K-Objectron subset.

## Usage

Each subset is loaded independently via `data_dir` (all share the same `train` split label — `data_dir` is what distinguishes them, not `split`):

```python
from datasets import load_dataset

ds_100K_Syn = load_dataset("pradhaansbhat/Thinking-In-Boxes", data_dir="data/100K-Syn", split="train")
ds_10K_Objectron = load_dataset("pradhaansbhat/Thinking-In-Boxes", data_dir="data/10K-Objectron", split="train")

print(ds_100K_Syn[0].keys())
# dict_keys(['bbox_0.png', 'bbox_1.png', 'rgb_0.png', 'rgb_1.png', '__key__', '__url__'])
```

Alternatively, using the named configs defined above:

```python
from datasets import load_dataset

ds_100K_Syn = load_dataset("pradhaansbhat/Thinking-In-Boxes", "100K-Syn", split="train")
ds_10K_Objectron = load_dataset("pradhaansbhat/Thinking-In-Boxes", "10K-Objectron", split="train")
ds_10K_Syn = load_dataset("pradhaansbhat/Thinking-In-Boxes", "10K-Objectron", split="train")
```

### Merging subsets for training

```python
from datasets import concatenate_datasets
from torch.utils.data import DataLoader

merged = concatenate_datasets([ds_10K_Objectron, ds_10K_Syn])  # e.g. combine the two 10K sets
loader = DataLoader(merged, batch_size=2, shuffle=True, num_workers=8)
```

**Note**: `__key__` values (e.g. `scene_000000`) are unique within each subset but **not** guaranteed unique across subsets after merging. This has no effect on training but is worth knowing if you rely on `__key__` for deduplication or lookups across merged data.

## Citation

If you find our work useful, please consider citing:

```bibtex
@misc{bhat2026thinkingboxes3dediting,
  title         = {Thinking in Boxes: 3D Editing in Real Images Made Easy},
  author        = {Pradhaan S Bhat and Naveen Chandra R and Rishubh Parihar and Vaibhav Vavilala and R. Venkatesh Babu and D. A. Forsyth and Anand Bhattad},
  year          = {2026},
  eprint        = {2606.20556},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.20556}
}
```
---