lrzjason/EditPair

Paired Image Editing Dataset

This dataset provides paired examples for training image editing models, containing original and modified images alongside text descriptions of the editing steps.

Dataset Structure

dataset_root/
├── {index}_F.png # Image F (e.g., edited version)
├── {index}_F.txt # Steps to transform G → F
├── {index}_G.png # Image G (e.g., original version)
└── {index}_G.txt # Steps to transform F → G

Naming Convention:

• Replace {index} with a digit identifier (e.g., 1, 2).

File Descriptions

• {index}_F.png & {index}_G.png: Paired images (e.g., before/after editing).
  • _F/G suffixes denote different states (e.g., F=edited, G=original).
• {index}_F.txt: Textual description of modifications applied to G.png to create F.png.
• {index}_G.txt: Textual description of modifications applied to F.png to revert to G.png.

Purpose

• Train models to understand/edit images through:
  • Image-to-image translation
  • Inpainting/outpainting
  • Style/attribute transfer
• Use text descriptions to guide conditional image generation.

Source & Editing

• Source Images: Collected from Unsplash (ensure compliance with Unsplash License).
• Edits: Applied using Jimeng mask edit techniques.

Contact

For questions or contributions, contact lrzjason at [email protected]