LiveCompose-outpainted / README_zh.md
zzsyppt's picture
Update Chinese README
11ff6a0 verified
metadata
pretty_name: Outpainted for Image Cropping
license: other
license_name: research-only
task_categories:
  - image-to-image
  - object-detection
tags:
  - image
  - computer-vision
  - image-cropping
  - bounding-box
  - outpainting
  - inpainting
  - stable-diffusion
  - composition
  - imagefolder
size_categories:
  - 10K<n<100K

Outpainted for Image Cropping

English | 中文

数据集简介

本数据集包含一批通过 Stable Diffusion v2 Inpaint 扩图得到的图像,以及每张扩图图像中“原始图像区域”的边界框标注。数据集主要面向图像裁剪、原始画幅恢复、构图感知裁剪和扩图感知视觉理解等研究任务。

每个样本包含:

  • 一张扩图后的图片;
  • orig_bbox:原图在扩展画布中的位置;
  • composition_tags:图片构图标签列表,部分为空。

数据生成流程

image-20260503132555963

数据生成流程如下:

  1. 收集专业摄影图片或高美学评分图片。
  2. 获取或生成图像描述,例如使用 BLIP 生成 caption。
  3. 设置待扩展边距。
  4. 使用 Stable Diffusion v2 Inpaint 对扩展区域进行补全。
  5. 使用正面提示词约束生成内容。
  6. 使用负面提示词减少不希望出现的内容,例如 framebordertextwatermark 等。
  7. 对生成结果进行伪影检测、一致性检测。
  8. 进行人工核查。
  9. 保留通过质量检查的样本,并记录原图区域 bbox,形成 training pairs。

orig_bbox 使用如下格式:

[x_min, y_min, x_max, y_max]

该 bbox 表示原始图像区域在扩图后画布中的位置,而不是目标检测中的物体框。

数据来源

本数据集的源图像来自或参考以下公开数据集/仓库:

  1. PICD: Photographic Image Composition Dataset
    https://github.com/CV-xueba/PICD_ImageComposition

  2. LAION Aesthetics v2 4.75
    https://huggingface.co/datasets/laion/aesthetics_v2_4.75

  3. Landscape-Dataset
    https://github.com/koishi70/Landscape-Dataset/tree/master

数据结构

outpainted-for-image-cropping/
├── README.md
├── metadata.jsonl
├── stats.json
└── images/
    ├── img_000000.png
    ├── img_000001.png
    └── ...

metadata.jsonl 中每一行对应一个样本,例如:

{
  "file_name": "images/img_000000.png",
  "orig_bbox": [281, 77, 881, 487],
  "composition_tags": ["HORI2"]
}

字段说明

  • file_name:扩图后图片的相对路径。
  • orig_bbox:原图区域在扩图画布中的边界框,格式为 [x_min, y_min, x_max, y_max]
  • composition_tags:从原始数据集中解析得到的构图标签列表。若没有可靠的构图标签,则为空列表 []

数据集统计

高频构图标签:

标签 数量
HORI2 1,956
HORI3 1,694
DIFFUSE 1,600
DENSE 1,436
DIA 1,305
LINE_VERTI3 1,156
PATTERN 1,000
LINE_VERTI_MANY 983
POINT_MULTI_HORI 64
LINE_VERTI2 55

使用方式

从 Hugging Face Hub 加载:

from datasets import load_dataset

dataset = load_dataset("zzsyppt/outpainted-for-image-cropping")
print(dataset)
print(dataset["train"][0])

本地上传前检查:

from datasets import load_dataset

dataset = load_dataset("imagefolder", data_dir="./hf_dataset")
print(dataset)
print(dataset["train"][0])

预期字段包括:

image
orig_bbox
composition_tags

引用

本数据集仅用于个人用途。如果使用本数据集,请根据实际使用的样本来源引用对应上游数据集。

PICD

@inproceedings{zhao2025can,
  title={Can Machines Understand Composition? Dataset and Benchmark for Photographic Image Composition Embedding and Understanding},
  author={Zhao, Zhaoran and Lu, Peng and Zhang, Anran and Li, Peipei and Li, Xia and Liu, Xuannan and Hu, Yang and Chen, Shiyi and Wang, Liwei and Guo, Wenhao},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={14411--14421},
  year={2025}
}

LAION-Aesthetics

请参考 LAION 官方页面和对应 Hugging Face 数据集页面引用 LAION-Aesthetics / LAION-5B 相关工作:

Landscape-Dataset

请参考原始仓库:

致谢

本数据集的生成使用了 Stable Diffusion v2 Inpaint,并参考或使用了公开图像数据源。感谢上游数据集、仓库和模型的创建者与维护者。