| --- |
| dataset_info: |
| features: |
| - name: url |
| dtype: string |
| - name: objects |
| list: |
| - name: bbox |
| list: int64 |
| - name: category |
| dtype: string |
| - name: long_local_prompt |
| dtype: string |
| - name: short_local_prompt |
| dtype: string |
| - name: long_global_caption |
| dtype: string |
| - name: short_global_caption |
| dtype: string |
| - name: image_hash |
| dtype: string |
| splits: |
| - name: train |
| num_bytes: 2672741454 |
| num_examples: 499249 |
| download_size: 665190298 |
| dataset_size: 2672741454 |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-* |
| --- |
| |
| # Abstract |
|
|
| Layout-to-Image generation has made substantial progress in image generation with spatial and object-level control. However, existing methods still struggle with complex scenes containing many overlapping and interacting objects. We argue that training data is a particular bottleneck: existing datasets lack examples with dense, complex object interactions. To address this gap, we introduce **OverLay++**, a large-scale Layout-to-Image dataset with structurally complex scenes. OverLay++ contains approximately 500K images with an average of 6.6 objects per image, exceeding existing datasets by 1.67× in annotation density. Beyond annotation density, OverLay++ provides rich semantic detail with object captions over six times longer than in current datasets. Our dataset generation pipeline is simple and robust, producing accurate overlapping regions with rich per-object captions. Across multiple benchmarks, state-of-the-art Layout-to-Image methods trained on OverLay++ dataset show consistent improvement and faster convergence demonstrating the importance of dense, overlap-aware, and caption-rich supervision for controllable image generation. Our code is available at [https://mlpc-ucsd.github.io/OverLayPP](https://mlpc-ucsd.github.io/OverLayPP). |
|
|
| ## Dataset description |
|
|
| Each example contains: |
|
|
| - `objects`: a list of annotated objects, each with: |
| - `bbox` |
| - `category` |
| - `long_local_prompt` |
| - `short_local_prompt` |
| - `long_global_caption` |
| - `short_global_caption` |
| - `url`: the original source URL |
| - `image_hash`: the original image identity hash when available |
|
|
| The object annotations pair image-coordinate bounding boxes with local prompts at two |
| levels of detail. Global captions summarize the complete image. |
|
|
| ## Example usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("mlpcucsd/OverLayPP", split="train", token=True) |
| sample = ds[0] |
| |
| print(sample["short_global_caption"]) |
| print(sample["objects"][0]["category"]) |
| print(sample["objects"][0]["short_local_prompt"]) |
| print(sample["url"]) |
| ``` |
|
|
