SEED-Data-Edit

SEED-Data-Edit is a hybrid dataset for instruction-guided image editing with a total of 3.7 image editing pairs, which comprises three distinct types of data:

Part-1: Large-scale high-quality editing data produced by automated pipelines (3.5M editing pairs).

Part-2: Real-world scenario data collected from the internet (52K editing pairs).

Part-3: High-precision multi-turn editing data annotated by humans (95K editing pairs, 21K multi-turn rounds with a maximum of 5 rounds).

This repo contains Part-1 of SEED-Data-Edit, with source images come from Unsplash.

After downloading the data, you first need to reassemble the split files back into the original .tar.gz file as below, and then unzip the files.

cat unsplash_images.tar.gz.part-* > unsplash_images.tar.gz

The folder "annotations" contains the original instructions, while this folder "annotations_GPT4V" stores a small portion of instructions that have been rewritten by GPT-4V.

SEED-X-Edit

You can download the image editing model SEED-X-Edit in Model, which is instruction tuned from the pre-trained SEED-X with SEED-Data-Edit.

For inference with SEED-X-Edit, you can refer to SEED-X.

License

SEED-Data-Edit is released under the license CC-BY-NC-4.0 for non-commercial research purpose only. Any use of the dataset for commercial purposes is strictly prohibited.

For Part-1, we use images from Unsplash and Openimages.

For Part-2, we collect images from Photoshopbattles, Photoshop gurus, Photoshoprequest, and Zhopped.

For Part-3, we use images from Unsplash, SAM, and JourneyDB.

Tencent does not hold the copyright for these images and the copyright belongs to the original owner.

If any image in SEED-Data-Edit infringes upon your rights, please contact us immediately and we will promptly remove the corresponding data.