PrismLayersPro

# PrismLayers: Open Data for High-Quality Multi-Layer Transparent Image Generative Models


We introduce **PrismLayersPro**, a 20K high-quality multi-layer transparent image dataset with rewritten style captions and human filtering.
**PrismLayersPro** is curated from our 200K dataset, **PrismLayers**, generated via MultiLayerFLUX.

### Dataset Structure

#### 📑 Dataset Splits (by Style)

The PrismLayersPro dataset is divided into 21 splits based on visual **style categories**.
Each split contains images sharing a consistent aesthetic (e.g., `toy`, `ink`, `anime`) and can be accessed individually via the `split` argument in the `load_dataset` function.

| Style | # Samples |
| ------------------- | --------- |
| 3D | 313 |
| Pokemon | 830 |
| anime | 717 |
| cartoon | 943 |
| doodle_art | 1065 |
| furry | 1038 |
| ink | 1155 |
| kid_crayon_drawing | 1001 |
| line_draw | 1221 |
| melting_gold | 946 |
| melting_silver | 1532 |
| metal_textured | 747 |
| neon_graffiti | 527 |
| papercut_art | 848 |
| pixel_art | 830 |
| pop_art | 782 |
| sand_painting | 929 |
| steampunk | 857 |
| toy | 1627 |
| watercolor_painting | 875 |
| wood_carving | 1263 |
| total | 20046 |

Below we show the structure of a single data sample from one style-specific split (e.g., `toy`).

Each image is composed of multiple RGBA layers, each with its own caption and layout information.

```
Example structure of a single sample (supports up to 50 layers):

{
"id": "sample_000001",
"style_category": "toy",
"whole_caption": "A colorful toy robot in a futuristic setting.",
"whole_image": <PIL.Image (RGBA)>,
"base_caption": "Futuristic neon-lit city background.",
"base_image": <PIL.Image (RGBA)>,
"layer_count": 6, # actual number of layers in this sample

# Repeated layer fields for up to 50 layers
"layer_00": <PIL.Image (RGBA)>,
"layer_00_caption": "Robot head with red visor",
"layer_00_box": [x, y, w, h],
"layer_00_width_dst": 256,
"layer_00_height_dst": 256,

"layer_01": <PIL.Image (RGBA)>,
"layer_01_caption": "Robot body with silver armor",
"layer_01_box": [x, y, w, h],
"layer_01_width_dst": 256,
"layer_01_height_dst": 256,

...

"layer_49": <PIL.Image (RGBA)>,
"layer_49_caption": "Shadow overlay with transparent gradient",
"layer_49_box": [x, y, w, h],
"layer_49_width_dst": 128,
"layer_49_height_dst": 128
}
```

### Dataset Stats

The figure below provides an overview of the dataset statistics.


## Load the dataset

<!-- Address questions around how the dataset is intended to be used. -->

To load the dataset, install the library `datasets` with `pip install datasets`. Then,

```
from datasets import load_dataset

dataset = load_dataset("artplus/PrismLayersPro", split="3D")
```

to download and load the style `3D`, for example.

# Citation

```text
@article{chen2025prismlayers,
title={PrismLayers: Open Data for High-Quality Multi-Layer Transparent Image Generative Models},
author={Chen, Junwen and Jiang, Heyang and Wang, Yanbin and Wu, Keming and Li, Ji and Zhang, Chao and Yanai, Keiji and Chen, Dong and Yuan, Yuhui},
journal={arXiv preprint arXiv:2505.22523},
year={2025}
}
```

# PrismLayers：面向高质量多层透明图像生成模型的开源数据集


我们推出了**PrismLayersPro**，这是一款包含2万条高质量多层透明图像的数据集，附带经过重写的风格化描述文本与人工筛选环节。**PrismLayersPro** 源自我们基于MultiLayerFLUX生成的20万条数据集**PrismLayers**。

### 数据集结构

#### 📑 按风格划分的数据集子集

PrismLayersPro 数据集基于视觉**风格类别**划分为21个子集。每个子集包含具有统一美学风格的图像（例如`toy`、`ink`、`anime`），可通过`load_dataset`函数中的`split`参数单独加载对应子集。

| 风格 | 样本数量 |
| ------------------- | --------- |
| 3D | 313 |
| 宝可梦（Pokemon） | 830 |
| 动画（anime） | 717 |
| 卡通（cartoon） | 943 |
| 涂鸦艺术（doodle_art） | 1065 |
| 兽人（furry） | 1038 |
| 水墨（ink） | 1155 |
| 儿童蜡笔画（kid_crayon_drawing） | 1001 |
| 线条画（line_draw） | 1221 |
| 熔金风格（melting_gold） | 946 |
| 熔银风格（melting_silver） | 1532 |
| 金属纹理（metal_textured） | 747 |
| 霓虹涂鸦（neon_graffiti） | 527 |
| 剪纸艺术（papercut_art） | 848 |
| 像素艺术（pixel_art） | 830 |
| 波普艺术（pop_art） | 782 |
| 沙画（sand_painting） | 929 |
| 蒸汽朋克（steampunk） | 857 |
| 玩具（toy） | 1627 |
| 水彩画（watercolor_painting） | 875 |
| 木雕（wood_carving） | 1263 |
| 总计 | 20046 |

下文将展示某一风格子集（例如`toy`）中单条数据样本的结构。每张图像由多个RGBA（红-绿-蓝-阿尔法通道）图层组成，每个图层均配有对应的描述文本与布局信息。

Example structure of a single sample (supports up to 50 layers):

{
"id": "sample_000001",
"style_category": "toy",
"whole_caption": "未来场景中的彩色玩具机器人。",
"whole_image": <PIL.Image (RGBA)>,
"base_caption": "未来霓虹城市背景。",
"base_image": <PIL.Image (RGBA)>,
"layer_count": 6, # 该样本的实际图层数量

# 最多支持50个图层的重复图层字段
"layer_00": <PIL.Image (RGBA)>,
"layer_00_caption": "带有红色护目镜的机器人头部。",
"layer_00_box": [x, y, w, h],
"layer_00_width_dst": 256,
"layer_00_height_dst": 256,

"layer_01": <PIL.Image (RGBA)>,
"layer_01_caption": "带有银色护甲的机器人躯体。",
"layer_01_box": [x, y, w, h],
"layer_01_width_dst": 256,
"layer_01_height_dst": 256,

...

"layer_49": <PIL.Image (RGBA)>,
"layer_49_caption": "带有透明渐变的阴影叠加层。",
"layer_49_box": [x, y, w, h],
"layer_49_width_dst": 128,
"layer_49_height_dst": 128
}

### 数据集统计信息

下图展示了该数据集的统计信息概览。


## 数据集加载方式

<!-- 解答有关数据集预期用途的相关问题。 -->

如需加载该数据集，请先通过`pip install datasets`安装`datasets`库，随后执行以下代码即可下载并加载对应风格的子集，例如：

from datasets import load_dataset

dataset = load_dataset("artplus/PrismLayersPro", split="3D")

## 引用格式

text
@article{chen2025prismlayers,
title={PrismLayers: Open Data for High-Quality Multi-Layer Transparent Image Generative Models},
author={Chen, Junwen and Jiang, Heyang and Wang, Yanbin and Wu, Keming and Li, Ji and Zhang, Chao and Yanai, Keiji and Chen, Dong and Yuan, Yuhui},
journal={arXiv preprint arXiv:2505.22523},
year={2025}
}