X2I Dataset

• Project Page: https://vectorspacelab.github.io/OmniGen/
• Github: https://github.com/VectorSpaceLab/OmniGen
• Paper: https://arxiv.org/abs/2409.11340
• Model: 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".

X2I-subject-driven

• Web-Image

A self-built subject-driven editing dataset with 36,316 & 45,425 & 111,734 samples.

## meta file: web-image-1.jsonl && web-image-2.jsonl && web-image-3.jsonl
cd retrieval
tar -zxvf download_images.tar.gz
tar -zxvf download_images_two.tar.gz

• GRIT-Entity

A subject-driven editing dataset with 1,708,742 samples.

## meta file: grit-entity.jsonl
cd grit/images1
tar -zxvf 00034.tar.gz
# tar -zxvf 00066.tar.gz 
# ...

cd grit/images2
tar -zxvf 00034.tar.gz
# tar -zxvf 00066.tar.gz 
# ...

cd grit/images3
tar -zxvf 00168.tar.gz
# tar -zxvf 00187.tar.gz 
# ...

• GRIT-Entity-New

A self-built subject-driven editing dataset with 676,603 samples. This datasets is smaller than GRIT-Entity but with higher qualtiy.

## meta file: grit-entity-new.jsonl
cd character
tar -xzvf character.tar.gz

cd human/human2
tar -xzvf human2.tar.gz

cd human/human3
tar -xzvf human3.tar.gz

cd single
cat single.tar.gz.* | tar -xzvf -

cd double
cat double.tar.gz.* | tar -xzvf -

cd triple
cat triple.tar.gz.* | tar -xzvf -