GPT-Image-Edit-1.5M: 1.5 Million Instruction-Guided Image Editing Pairs

GPT-Image-Edit-1.5M consists of 1.5 million image editing pairs produced by UCSC-VLAA in 2025. The collection aggregates samples from HQ-Edit, UltraEdit, OmniEdit, and Complex-Edit, with all output images systematically regenerated using the GPT-Image-1 model.

Use Cases

• Instruction-guided image editing using the text prompts and image pairs
• Instruction-tuning for multimodal models on image-to-image tasks
• Benchmarking generative models on complex editing instructions derived from OmniEdit

Strengths

• 1.5 million image pairs
• CC BY 4.0 license
• Aggregates four major editing datasets including HQ-Edit and OmniEdit
• Unified output quality via GPT-Image-1 regeneration

Limitations

• Synthetic output images may contain artifacts inherent to the GPT-Image-1 generator
• Inherits any geographic or topical biases present in the four source datasets

Provenance

Source

UCSC-VLAA; Arxiv 2507.21033

Collection Method

Synthetic regeneration of existing datasets (HQ-Edit, UltraEdit, OmniEdit, Complex-Edit) using GPT-Image-1.

Time Range

2025

Freshness

Last updated August 2025; based on research published in 2025.