MultiEdit Multimodal Image Editing Dataset

MultiEdit is a comprehensive, large-scale instruction-based image editing dataset released in 2025 by inclusionAI in collaboration with the University of New South Wales and the University of Hong Kong.MultiEdit: Advancing Instruction-based Image Editing on Diverse and Challenging Tasks", which aims to improve the model's capabilities in complex and diverse image editing tasks.

This dataset contains approximately 107,000 samples, covering six major editing tasks and 56 subcategories of edit types, including object reference editing, person reference editing, text and interface element adjustment, perspective transformation, and style transfer. The data is derived from a generation process driven by large multimodal models (such as GPT-4o and GPT-Image-1). This approach combines instruction construction, image generation, and quality screening to ensure the relevance and consistency of the edit samples. The data structure consists of a triplet of "source image - edit instruction - edit result," along with information about the edit category and source.

Data composition

• Object Reference Editing
  • Used to modify the properties of specific objects, including color, shape, scale, and position.
  • It contains 4 editing types and a total of 10,051 samples (9,851 in the training set and 200 in the test set).
• Person Reference Editing
  • Edit the people in the image, including posture, clothing, hairstyle, skin color, and body shape.
  • It contains 5 types of edits, with a total of 7,141 samples (6,891 in the training set and 250 in the test set).
• Text Editing
  • Modify text elements in images, such as font style, text content, display medium, and color.
  • It contains 4 types of edits, with a total of 4,060 samples (3,860 in the training set and 200 in the test set).
• GUI Editing
  • Used to edit the icon properties and display media of graphical user interface (GUI) elements, covering iOS, Android and web interfaces.
  • It contains 2 types of edits, with a total of 2,880 samples (2,780 in the training set and 100 in the test set).
• View Editing
  • Generate different perspectives of image subjects, including people, landmarks, and general objects.
  • It contains 3 types of edits, with a total of 28,205 samples (28,055 in the training set and 150 in the test set).
• Style Transfer
  • Convert images into 38 art styles, from classic art forms to modern digital aesthetics.
  • It contains 38 editing types and a total of 56,297 samples (55,097 in the training set and 200 in the test set).