X2I-text-to-image

# X2I Dataset

* Project Page: [https://vectorspacelab.github.io/OmniGen/](https://vectorspacelab.github.io/OmniGen/)
* Github: [https://github.com/VectorSpaceLab/OmniGen](https://github.com/VectorSpaceLab/OmniGen)
* Paper: [https://arxiv.org/abs/2409.11340](https://arxiv.org/abs/2409.11340)
* Model: [https://huggingface.co/Shitao/OmniGen-v1](https://huggingface.co/Shitao/OmniGen-v1)

To achieve robust multi-task processing capabilities, it is essential to train the **OmniGen** on large-scale and diverse datasets. However, in the field of unified image generation, a readily available dataset has yet to emerge. For this reason, we have curated a large-scale **unified image generation** dataset with unified format for the **first time**, which we refer to as the **X2I dataset**, meaning **"anything to image"**.

| Task| Datastet|
| :-------- | :-------- |
| Multi-modal Instruction| [X2I-mm-instruction](https://huggingface.co/datasets/yzwang/X2I-mm-instruction) |
| Subject-driven Editing | [X2I-subject-driven](https://huggingface.co/datasets/yzwang/X2I-subject-driven) |
| In-context Learning | [X2I-in-context-learning](https://huggingface.co/datasets/yzwang/X2I-in-context-learning) |
| Computer Vision | [X2I-computer-vision](https://huggingface.co/datasets/yzwang/X2I-computer-vision) |
| Text to Image Generation| [X2I-text-to-image](https://huggingface.co/datasets/yzwang/X2I-text-to-image) |

## X2I-text-to-image

- **laion-coco-aesthetic**

A subset of [LAION-COCO](https://huggingface.co/datasets/laion/laion-coco) with 4,134,263 images filtered (image size > 384x384, aesthetic score > 4.75, watermark probability < 0.5) as [this](https://huggingface.co/datasets/guangyil/laion-coco-aesthetic).

```python
## meta file: laion-coco-aesthetic.jsonl
cd laion-coco-aesthetic
tar -xzvf 00000.tar.gz
# tar -xzvf 00001.tar.gz
# tar -xzvf 00002.tar.gz
# ...
```

- **other open-source datasets**
- [Recap-DataComp-1B](https://huggingface.co/datasets/UCSC-VLAA/Recap-DataComp-1B)
- [SAM-LLaVA-Captions10M](https://huggingface.co/datasets/PixArt-alpha/SAM-LLaVA-Captions10M)
- [ALLaVA-4V](https://huggingface.co/datasets/FreedomIntelligence/ALLaVA-4V)
- [DOCCI](https://huggingface.co/datasets/google/docci)
- [ShareGPT4V](https://huggingface.co/datasets/Lin-Chen/ShareGPT4V)
- [DenseFusion-1M](https://huggingface.co/datasets/BAAI/DenseFusion-1M)
- [JourneyDB](https://huggingface.co/datasets/JourneyDB/JourneyDB)

# X2I 数据集（X2I Dataset）
* 项目页面：[https://vectorspacelab.github.io/OmniGen/](https://vectorspacelab.github.io/OmniGen/)
* Github：[https://github.com/VectorSpaceLab/OmniGen](https://github.com/VectorSpaceLab/OmniGen)
* 论文：[https://arxiv.org/abs/2409.11340](https://arxiv.org/abs/2409.11340)
* 模型：[https://huggingface.co/Shitao/OmniGen-v1](https://huggingface.co/Shitao/OmniGen-v1)

为实现鲁棒的多任务处理能力，需在大规模且多样化的数据集上对**OmniGen**进行训练。然而，当前统一图像生成领域尚未有成熟可用的公开数据集。为此，我们首次构建了格式统一的大规模**统一图像生成**数据集，将其命名为**X2I 数据集**，意为「万物转图像（anything to image）」。

| 任务类型 | 数据集 |
| :-------- | :-------- |
| 多模态指令 | [X2I-mm-instruction](https://huggingface.co/datasets/yzwang/X2I-mm-instruction) |
| 主体驱动编辑 | [X2I-subject-driven](https://huggingface.co/datasets/yzwang/X2I-subject-driven) |
| 上下文学习 | [X2I-in-context-learning](https://huggingface.co/datasets/yzwang/X2I-in-context-learning) |
| 计算机视觉 | [X2I-computer-vision](https://huggingface.co/datasets/yzwang/X2I-computer-vision) |
| 文本到图像生成 | [X2I-text-to-image](https://huggingface.co/datasets/yzwang/X2I-text-to-image) |

## X2I-text-to-image
- **laion-coco-aesthetic**
该数据集为 [LAION-COCO](https://huggingface.co/datasets/laion/laion-coco) 的子集，共包含4,134,263张经过筛选的图像（筛选条件为图像分辨率>384×384、美学评分>4.75、水印概率<0.5），具体可参见 [该数据集页面](https://huggingface.co/datasets/guangyil/laion-coco-aesthetic)。
python
## 元数据文件：laion-coco-aesthetic.jsonl
cd laion-coco-aesthetic
tar -xzvf 00000.tar.gz
# tar -xzvf 00001.tar.gz
# tar -xzvf 00002.tar.gz
# ...

- **其他开源数据集**
- [Recap-DataComp-1B](https://huggingface.co/datasets/UCSC-VLAA/Recap-DataComp-1B)
- [SAM-LLaVA-Captions10M](https://huggingface.co/datasets/PixArt-alpha/SAM-LLaVA-Captions10M)
- [ALLaVA-4V](https://huggingface.co/datasets/FreedomIntelligence/ALLaVA-4V)
- [DOCCI](https://huggingface.co/datasets/google/docci)
- [ShareGPT4V](https://huggingface.co/datasets/Lin-Chen/ShareGPT4V)
- [DenseFusion-1M](https://huggingface.co/datasets/BAAI/DenseFusion-1M)
- [JourneyDB](https://huggingface.co/datasets/JourneyDB/JourneyDB)