KempnerInstituteAI/flux.2-dev-synthetic-2M

---
license: other
language:
- en

task_categories:
- text-to-image

annotations_creators:
- machine-generated

source_datasets:
- text-to-image-2M

tags:
- synthetic
- diffusion
- text-to-image
- webdataset
- flux
- generative-models

size_categories:
- 1M<n<10M

viewer: false

configs:
- config_name: default
data_files:
- split: train
path: train/*.tar
---

# Flux2.dev Synthetic: 2.2M Text-to-Image Pairs at 512×512

## Dataset Summary

This dataset contains **~2.2 million** large-scale synthetic image–caption pairs generated using the FLUX.2-dev diffusion model:

- **Model**: [black-forest-labs/FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev)
- **Caption source**: [Text2Image-2M](https://huggingface.co/datasets/jackyhate/text-to-image-2M)
- **Total samples**: 2,282,665 (571 shards × ~4000 samples)
- **Resolution**: 512 × 512
- **Image format**: PNG (lossless)
- **Total shards**: 571
- **Samples per shard**: 4000 (last shard: 2665)
- **Total size**: ~865 GB

Each sample consists of:
- `.png` image file
- `.txt` caption file (with the same base name as the image)
- `.json` metadata file (with the same base name as the image)

The dataset is stored in WebDataset format for scalable streaming and distributed training.

## Example Samples

| Caption | Image |
|---------|-------|
| A woman in a hat stands in front of a colorful umbrella. |  |
| A plate of Chinese food with a variety of dishes including dumplings, noodles, and vegetables. There is a small bowl of red sauce on the side. The plate is placed on a white tablecloth, and there are chopsticks resting on the tablecloth to the right of the plate. |  |
| an open air sculpture with trees in the background |  |
| An animated figure, likely a young girl with green eyes and black hair, enjoys a meal of noodles at an outdoor wooden table amidst a calm, scenic environment. |  |

## Dataset Statistics

| Property | Value |
|--------|------|
| Total Samples | 2,282,665 |
| Total Shards | 571 |
| Samples per Shard | 4000 (last shard: 2665) |
| Image Resolution | 512×512 |
| Image Format | PNG |
| Total Dataset Size | ~865 GB |
| Caption Source | Text2Image-2M |
| Generation Model | FLUX.2-dev |
| Inference Steps | 50 |
| Guidance Scale | 3.5 |

## Dataset Structure

```bash
flux2_synthetic_000000.tar
flux2_synthetic_000001.tar
...
```

Each `.tar` shard contains grouped samples:

```bash
00020999.png
00020999.txt
00020999.json
```

## Example Metadata

```json
{
"id": 20999,
"seed": 42,
"num_inference_steps": 50,
"guidance_scale": 3.5,
"height": 512,
"width": 512,
"model": "black-forest-labs/FLUX.2-dev",
"shard_id": 5
}
```

## Intended Use

The captions are sourced from the Text2Image-2M dataset, which contains a wide variety of image descriptions. The synthetic images were generated using fixed parameters (50 steps, guidance scale 3.5) to create a large-scale dataset for training and evaluating text-to-image models. Any use of this dataset should follow the terms of the original FLUX.2-dev and Text2Image-2M licenses. See the License section below for details.

## Loading the Dataset

### Using WebDataset (Recommended for PyTorch)

```python
import webdataset as wds
import json

dataset = (
wds.WebDataset("path/to/flux2_synthetic_{000000..000570}.tar")
.decode("pil")
.to_tuple("png", "txt", "json")
)

for image, caption, metadata in dataset:
# image: PIL Image
# caption: str
# metadata: dict (automatically parsed from JSON)
print(f"Caption: {caption}")
print(f"Seed: {metadata['seed']}")
```

### Streaming from Hugging Face (Example)

```python
from datasets import load_dataset

dataset = load_dataset(
"webdataset"
data_files="https://huggingface.co/datasets/KempnerInstituteAI/flux.2-dev-synthetic-2M/resolve/main/{000000..000570}.tar",
split="train",
streaming=True
)

# Example: iterate through samples
for sample in dataset:
image = sample["image"]
caption = sample["text"]
metadata = sample["json"]
# Your processing here
```

## General Details

Images were generated using:

- Model: black-forest-labs/FLUX.2-dev
- Steps: 50 inference steps
- Guidance scale: 3.5
- Seed: Stored per-sample
- Resolution: 512×512
- Format: PNG (lossless)

Captions originate from the Text2Image-2M corpus.

## Dataset Creation

This dataset was generated using a distributed pipeline:

1. Captions were extracted from the Text2Image-2M dataset
2. Images were generated using FLUX.2-dev with fixed parameters (50 steps, guidance 3.5)
3. Generation was distributed across 571 shards for parallel processing
4. Each sample includes the original caption and generation metadata
5. Output was packaged in WebDataset format for efficient streaming

## Data Characteristics

- **All images are synthetic** (generated by FLUX.2-dev, not real photographs)
- Captions were not modified from the Text2Image-2M source
- No additional filtering applied beyond successful generation
- No watermarking was applied
- No human annotations were added
- All samples use consistent generation parameters (steps=50, guidance=3.5)

## Limitations and Considerations

- **Synthetic Biases**: Images reflect the biases and characteristics of the FLUX.2-dev model
- **Caption Quality**: Caption quality depends entirely on the Text2Image-2M dataset
- **Limited Diversity**: Generated with fixed parameters, which may limit stylistic diversity
- **No Quality Filtering**: All successfully generated images are included without quality assessment
- **Generation Artifacts**: A small number of samples may contain visual artifacts or generation failures typical of diffusion-based image synthesis
- **Not Ground Truth**: Synthetic images should not be treated as factually accurate representations

## License

Users are responsible for ensuring compliance with all applicable licenses.

This dataset combines content from two sources, each with its own license:

### Images: FLUX.2-dev License

All images are synthetic outputs generated using [black-forest-labs/FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev).

- **You MUST review the [FLUX.2-dev license](https://huggingface.co/black-forest-labs/FLUX.2-dev/blob/main/LICENSE.md)** before using this dataset.

### Captions: Text2Image-2M License

Captions originate from the [Text2Image-2M](https://huggingface.co/datasets/jackyhate/text-to-image-2M) dataset.

- **Review the [Text2Image-2M license](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/mit.md)** for caption usage terms.

### This Dataset

This dataset compilation is intended for research purposes only. Any use should proceed only after confirming compliance with the applicable upstream licenses.

## Citation

If you use this dataset, please cite:

```bibtex
@misc{naeem_khoshnevis_2026,
author = { Naeem Khoshnevis and Gabriel Guo and Eric Vanden-Eijnden and Nicholas Boffi and Michael S. Albergo },
title = { flux.2-dev-synthetic-2M (Revision 0e2762f) },
year = { 2026 },
note = { Generated with FLUX.2-dev using Text2Image-2M captions. The project was co-advised by Nicholas Boffi and Michael S. Albergo. },
url = { https://huggingface.co/datasets/KempnerInstituteAI/flux.2-dev-synthetic-2M },
doi = { 10.57967/hf/8311 },
publisher = { Hugging Face }
}
```

Additionally, please cite the FLUX.2-dev model:

```bibtex
@software{flux2_dev,
title={FLUX.2-dev},
author={Black Forest Labs},
year={2024},
url={https://huggingface.co/black-forest-labs/FLUX.2-dev}
}
```

And the Text2Image-2M dataset:

```bibtex
@misc{zk_2024,
author = { zk },
title = { text-to-image-2M (Revision e64fca4) },
year = 2024,
url = { https://huggingface.co/datasets/jackyhate/text-to-image-2M },
doi = { 10.57967/hf/3066 },
publisher = { Hugging Face }
}
```

## Acknowledgements

This dataset was generated as part of research conducted by Naeem Khoshnevis, Gabriel Guo, Eric Vanden-Eijnden, Nicholas Boffi, and Michael S. Albergo, with Nicholas Boffi and Michael S. Albergo serving as shared advisors on the project. The work used compute resources provided by the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University.

The large-scale image generation pipeline relied on the Institute's high-performance GPU infrastructure, including distributed compute clusters and high-throughput storage systems. We gratefully acknowledge the technical support and infrastructure engineering efforts that made this large-scale synthetic data generation possible.