---
license: cc-by-nc-4.0
task_categories:
  - image-to-image
tags:
  - image-editing
  - camera-control
  - focal-plane
  - depth-of-field
  - diffusion-model
  - photography
pretty_name: CamEdit50K - Focal Plane
size_categories:
  - 10K<n<100K
---

# CamEdit50K: Focal Plane Subset

This dataset is part of the **CamEdit** project — a diffusion-based framework for continuous camera parameter control in photorealistic image editing. This subset focuses on **focal plane** manipulation, providing paired data for training models that can re-render images at different focal planes via text instructions.

Paper: *CamEdit: Continuous Camera Parameter Control for Photorealistic Image Editing*

## Dataset Summary

| Statistic | Value |
|---|---|
| Total image pairs | 30,703 |
| Unique input images | 19,562 |
| Unique edited images | 29,790 |
| Focal plane parameter range | [0.0, 1.0] |
| Image resolution | 540–2,604 px (mean ~901 x 1,274) |
| Format | JPEG |
| Number of shards | 11 |

## Dataset Structure

### Files

Each shard consists of a `.tar` archive containing the images and a `.parquet` file containing the metadata:

```
CamEdit_Focal_001.tar        # images (input/ and gt/ subfolders)
CamEdit_Focal_001.parquet    # metadata for this shard
...
CamEdit_Focal_011.tar
CamEdit_Focal_011.parquet
CamEditckpt_@_.tar           # pretrained LoRA checkpoints (AP / FP / SP)
```

### Image Folder Layout

Each tar archive extracts to:

```
CamEdit_Focal_XXX/
├── input/          # original images
│   ├── 0000049.jpg
│   └── ...
└── gt/             # ground-truth edited images
    ├── 0000049-focal-0.3.jpg
    └── ...
```

### Parquet Schema

| Column | Type | Description |
|---|---|---|
| `index` | string | Unique sample ID |
| `original_image_path` | string | Relative path to the input image |
| `edited_image_path` | string | Relative path to the ground-truth edited image |
| `instruction` | string | Text instruction, e.g. *"Render the image with focal plane 0.3."* |
| `height` | int | Image height in pixels |
| `width` | int | Image width in pixels |
| `task` | string | Task type (`Focal Plane`); absent in shards 008–011 |
| `parameter` | string | Focal plane value (0–1) |

### Per-Shard Statistics

| Shard | Input Images | Pairs | Size |
|---|---|---|---|
| Focal_001 | 1,942 | 3,000 | 3.75 GB |
| Focal_002 | 1,941 | 3,000 | 3.70 GB |
| Focal_003 | 1,935 | 3,000 | 3.72 GB |
| Focal_004 | 1,961 | 3,000 | 3.79 GB |
| Focal_005 | 1,940 | 3,000 | 3.72 GB |
| Focal_006 | 1,936 | 3,000 | 3.66 GB |
| Focal_007 | 1,626 | 2,620 | 2.87 GB |
| Focal_008 | 530 | 3,000 | 1.17 GB |
| Focal_009 | 1,668 | 3,000 | 1.27 GB |
| Focal_010 | 3,000 | 3,000 | 2.16 GB |
| Focal_011 | 1,083 | 1,083 | 0.81 GB |

### Checkpoints

`CamEditckpt_@_.tar` contains pretrained LoRA weights and tokenizer embeddings for three camera parameter editing tasks:

- **AP** — Aperture
- **FP** — Focal Plane
- **SP** — Shutter Speed

Each subfolder includes:
- `pytorch_lora_weights.safetensors`
- `Embed_one/`, `Embed_two/`, `Embed_three/` (learned camera token embeddings)
- `tokenizer/`, `tokenizer_2/`, `tokenizer_3/`

## Usage

```python
import pandas as pd, tarfile

# Read metadata
df = pd.read_parquet("CamEdit_Focal_001.parquet")

# Extract images
with tarfile.open("CamEdit_Focal_001.tar") as tar:
    tar.extractall(".")

# Access a sample
row = df.iloc[0]
print(row.instruction)        # "Render the image with focal plane 0.3."
print(row.original_image_path) # "CamEdit_Focal_001/input/0042680.jpg"
print(row.edited_image_path)   # "CamEdit_Focal_001/gt/0042680-focal-0.3.jpg"
```

## License

This dataset is released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).

## Citation

If you use this dataset, please cite:

```bibtex
@article{qin2026camedit,
  title={Camedit: Continuous camera parameter control for photorealistic image editing},
  author={Qin, Xinran and Wang, Zhixin and Li, Fan and Chen, Haoyu and Pei, Renjing and Li, Wenbo and Cao, Xiaochun},
  journal={Advances in Neural Information Processing Systems},
  volume={38},
  pages={114152--114171},
  year={2026}
}
```