README.md

---
license: other
license_name: mixed-per-image-licenses
license_link: LICENSE
pretty_name: Breakpoint Grounding 55M
task_categories:
- object-detection
- text-to-image
- image-to-text
language:
- en
tags:
- object-detection
- bounding-boxes
- grounding
- image-captioning
- synthetic-data
- diffusion-models
size_categories:
- 10M<n<100M
dataset_info:
  features:
  - name: image
    dtype: image
  - name: image_caption
    dtype: string
  - name: object_captions
    sequence: string
  - name: normalized_boxes
    sequence:
      sequence: float32
  - name: img_size_wh
    sequence: int32
  splits:
  - name: train
    num_examples: 54909477
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# Breakpoint Grounding 55M

## Quick start

```python
from datasets import load_dataset
ds = load_dataset("BreakpointAI/breakpoint-grounding-55m", split="train")
ds[0]  # {'image': <PIL.Image>, 'image_caption': ..., 'object_captions': [...], 'normalized_boxes': [...], 'img_size_wh': [...]}
```

## Dataset summary

**Breakpoint Grounding 55M is, to our knowledge, the largest instance-grounded image–text dataset
released publicly.** Every image comes with an **image-level caption**, a set of **object bounding
boxes**, and **a caption for every box**. It was built by [Breakpoint AI](https://huggingface.co/BreakpointAI)
to train a 10B-parameter joint image + bounding-box diffusion model for synthetic object-detection
data generation.

![An image with bounding boxes drawn over detected objects, each box captioned, alongside the
image's global caption](./example_figure.png)

- **54,909,477 images**, each with one image-level caption
- **511,652,664 bounding boxes** (9.32 per image on average), each with its own caption

For comparison, **ROVI** (ICCV 2025), the most recent comparable instance-grounded dataset, labels
1M web images with a similar VLM + open-vocabulary-detector pipeline — this dataset is more than
**50× larger**.

## Why this dataset exists

At Breakpoint, we trained diffusion models to generate synthetic training data for object
detection models. Our ultimate model was a 10-billion-parameter joint diffusion model that
generated images and bounding boxes simultaneously. Given 10 labeled images, it could adapt to a
new scene and start producing labeled training images.

In academia, there has been considerable research on conditional generation (bounding boxes to
image) (e.g. GLIGEN [Li et al., 2023], InstanceDiffusion [Wang et al., 2024]), and the associated
datasets (e.g. ROVI [Peng et al., 2025]). While helpful, this work did not touch joint generation,
and unfortunately was not at the scale necessary to be actually useful.

In order to get a better performing model, we decided to build this dataset, which is over 50
times larger than comparable datasets (e.g. ROVI, ICCV 2025, at 1M images). Having this much data
is what allowed us to train a large and high-performing model.

## Image sources

Images are drawn from four public web-image collections:

| Source | Original image licensing | Notes |
|--------|--------------------------|-------|
| [Open Images](https://storage.googleapis.com/openimages/web/index.html) | CC BY 2.0 (images), CC BY 4.0 (Google annotations) | Attribution per the Open Images terms. |
| [Wikimedia Commons](https://commons.wikimedia.org/) | Mixed free licenses (CC BY-SA, CC BY, CC0, public domain), per file | Attribution and license vary per file. |
| [RedCaps](https://redcaps.xyz/) | Reddit-submitted images; released for non-commercial research under the RedCaps terms | Subject to the RedCaps takedown process. |
| [LAION](https://laion.ai/) | URL/metadata under CC BY 4.0; images remain under their original owners' rights | Image bytes are redistributed here rather than URLs — see Licensing. |

## How it was built

- **Images**: collected from the four sources above, resized to ~1MP, light integrity filtering
  (corrupt / unreadable files removed).
- **Annotations** (`image_caption`, `object_captions`, `normalized_boxes`): generated by
  Breakpoint AI's automated annotation pipeline (object detection + captioning models). They are
  **not human-labeled**. See **Appendix: Annotation pipeline** for details.

## Dataset structure

### Fields

| Field | Type | Description |
|-------|------|-------------|
| `image` | `Image` | The decoded image (embedded in the Parquet files). |
| `image_caption` | `string` | A single natural-language caption describing the whole image. |
| `object_captions` | `Sequence(string)` | One caption per detected object. Index-aligned with `normalized_boxes`. |
| `normalized_boxes` | `Sequence(Sequence(float32))` | One bounding box per detected object, coordinates normalized to `[0, 1]`. Index-aligned with `object_captions`. |
| `img_size_wh` | `Sequence(int32)` | Original image size as `[width, height]` in pixels. |

Which of the four source datasets each row came from is published separately, as
`metadata/source_dataset_index.parquet` — see
[Which rows come from which source](#which-rows-come-from-which-source).

### Bounding box format

Each box is `[x_min, y_min, x_max, y_max]` (xyxy), with coordinates normalized to `[0, 1]`
relative to image width/height. `normalized_boxes[i]` corresponds to `object_captions[i]`.

### Splits

| Split | Rows |
|-------|------|
| `train` | 54,909,477 |

### Example

```python
from datasets import load_dataset

ds = load_dataset("BreakpointAI/breakpoint-grounding-55m", split="train", streaming=True)
row = next(iter(ds))

row["image"]              # PIL.Image
row["image_caption"]      # "a photograph of ..."
row["object_captions"]    # ["a dog", "a red ball", ...]
row["normalized_boxes"]   # [[0.12, 0.34, 0.56, 0.78], ...]
row["img_size_wh"]        # [1820, 1024]
```

## Licensing and responsible use

This dataset aggregates images from multiple public sources, each under its own terms. They fall
into three groups, not one blanket restriction:

- **No research-only restriction — usable for any purpose, including commercial, per the
  source's own license (with attribution where required):**
  - The bounding boxes and captions contributed by Breakpoint AI itself — **CC BY 4.0**.
  - Images from **Open Images** — CC BY 2.0 (images), CC BY 4.0 (Google's annotations).
  - Images from **Wikimedia Commons** — CC BY-SA, CC BY, CC0, or public domain, per file; check
    the specific file's license for attribution/share-alike requirements.
- **Restricted to non-commercial research use:**
  - Images from **RedCaps** — under the RedCaps terms, non-commercial research only.
- **Rights not cleared by Breakpoint — neither confirmed open nor research-restricted, genuinely
  case-by-case:**
  - Images from **LAION** — the URLs/metadata are CC BY 4.0, but each image remains under its
    original, third-party owner's rights, which Breakpoint has not individually verified.
    Downstream users are responsible for determining the applicable rights before using these
    images beyond research.

### Which rows come from which source

| Source | Rows | Share | Licensing group (above) |
|--------|-----:|------:|-------------------------|
| LAION | 37,392,932 | 68.10% | Rights not cleared — case-by-case |
| Wikimedia Commons | 7,344,982 | 13.38% | No research-only restriction |
| RedCaps | 5,290,815 | 9.64% | Non-commercial research only |
| Open Images | 4,880,748 | 8.89% | No research-only restriction |

The rows themselves don't carry a source column, but **`metadata/source_dataset_index.parquet`**
in this repo maps every row to its source. It has one row per dataset row, in the same order as
the `train` split, with columns `index` (int32) and `source_dataset` (one of `laion`,
`wikimedia`, `redcaps`, `openimages`).

```python
from datasets import load_dataset
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

ds = load_dataset("BreakpointAI/breakpoint-grounding-55m", split="train")

path = hf_hub_download(
    repo_id="BreakpointAI/breakpoint-grounding-55m",
    repo_type="dataset",
    filename="metadata/source_dataset_index.parquet",
)
sources = pq.read_table(path).column("source_dataset").to_pylist()

ds = ds.add_column("source_dataset", sources)          # now filterable
redcaps_only = ds.filter(lambda r: r["source_dataset"] == "redcaps")
```

The join is positional, so it only holds if you haven't shuffled or filtered the dataset before
attaching the column.

Use of the images is governed by the applicable source's terms above; downstream users are
responsible for complying with them.

**Personal data.** The images depict real, sometimes identifiable, people and places. Material
being publicly posted does not remove it from the scope of data-protection laws such as the GDPR,
UK GDPR, and CCPA. Do **not** use this dataset to identify, profile, track, surveil, or contact
individuals, or to train biometric identification systems.

**No affiliation.** This dataset is not affiliated with, endorsed by, or connected to Google /
Open Images, the Wikimedia Foundation, Reddit, or LAION e.V.

**Removal requests.** If you are the rights holder for an image, or are depicted in one, and want
it removed, open a discussion on this repository identifying the affected row indices (or attaching
the image) and the affected rows will be removed.

## Limitations and biases

- Annotations are model-generated and contain errors: missed objects, spurious boxes, imprecise
  coordinates, and caption hallucinations.
- Inherits the content distribution and biases of the four source collections, including
  web-scale skews in geography, language, subject matter, and the demographics of who posts
  images to Reddit and Wikimedia.
- Images vary in aesthetic and resolution quality; only a light integrity filter was applied.

## Acknowledgements

- Hosted on the Hugging Face Hub with a public dataset storage grant.
- Thanks to Daniel van Strien and the Hugging Face datasets team.
- Images from Open Images, Wikimedia Commons, RedCaps, and LAION.

## Appendix: Annotation pipeline

- **Object detection**: bounding boxes and class labels were produced by an open-vocabulary
  detector (Grounding DINO) as an upstream step.
- **Per-object captions**: [InternVL3-2B](https://huggingface.co/OpenGVLab/InternVL3-2B), served
  via LMDeploy (TurboMind engine, `internvl2_5` chat template). Each detected box was cropped
  (no padding) and captioned with the prompt:
  > `<image>` Describe this `<category name>` in 20 words or less and focus on its color,
  > appearance, size, and style, etc.

  where `<category name>` is the detector's class label for that box. Generated with
  `max_new_tokens=64`.
- **Image-level captions**: [InternVL3-8B](https://huggingface.co/OpenGVLab/InternVL3-8B) (via
  `transformers`), prompted with:
  > `<image>` Write a single, natural-language description of this image. Do not use bullet
  > points or lists. Keep it under 100 words.

  Generated with `max_new_tokens=128`, sampling enabled.

## Citation

```bibtex
@misc{breakpoint_grounding_55m,
  title  = {Breakpoint Grounding 55M},
  author = {Wang, Franklin and Zhong, Desmond and Murdoch, Jamie},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BreakpointAI/breakpoint-grounding-55m}
}
```

### Related work

```bibtex
@inproceedings{peng2025rovi,
  title     = {ROVI: A VLM-LLM Re-Captioned Dataset for Open-Vocabulary
               Instance-Grounded Text-to-Image Generation},
  author    = {Peng, Cihang and Hou, Qiming and Ren, Zhong and Zhou, Kun},
  booktitle = {ICCV},
  year      = {2025}
}

@inproceedings{li2023gligen,
  title     = {GLIGEN: Open-Set Grounded Text-to-Image Generation},
  author    = {Li, Yuheng and Liu, Haotian and Wu, Qingyang and Mu, Fangzhou and Yang, Jianwei
               and Gao, Jianfeng and Li, Chunyuan and Lee, Yong Jae},
  booktitle = {CVPR},
  year      = {2023}
}

@inproceedings{wang2024instancediffusion,
  title     = {InstanceDiffusion: Instance-level Control for Image Generation},
  author    = {Wang, Xudong and Darrell, Trevor and Rambhatla, Sai Saketh and Girdhar, Rohit
               and Misra, Ishan},
  booktitle = {CVPR},
  year      = {2024}
}
```

<!-- TODO: verify these two entries (author order / exact venue) before publishing — drafted
     from memory, not looked up. -->