OpenGPT-4o-Image: A Comprehensive Dataset for Advanced Image Generation and Editing
| Source: arXiv AI
Tags: image generation, image editing, multimodal, dataset, GPT-4o, OpenGPT-4o-Image, fine-tuning
Researchers release OpenGPT-4o-Image, an 80K instruction-image pair dataset spanning 11 domains and 51 subtasks including chemistry diagrams and complex multi-step editing — fine-tuning on it improves image editing benchmarks by up to 18% and generation by 13%.
Details
Current multimodal image generation/editing datasets cover basic tasks but lack systematic coverage of challenging scenarios. OpenGPT-4o-Image addresses this with a hierarchical taxonomy — ranging from fundamental capabilities (text rendering, style control) to highly specialized tasks (scientific imagery for chemistry, complex instruction editing requiring simultaneous operations).\n\nThe dataset was built using an automated pipeline combining structured resource pools with GPT-4o, yielding 80K high-quality instruction-image pairs. The 51-subtask taxonomy across 11 major domains makes it one of the most systematically structured public datasets in this space.\n\nFine-tuning experiments show significant performance gains: UniWorld-V1 improves 18% on ImgEdit-Bench editing tasks and Harmon achieves 13% improvement on GenEval generation benchmarks. These are meaningful gains from a publicly released dataset.\n\nFor practitioners training or fine-tuning image generation models, this fills a specific gap: complex, multi-operation editing scenarios and scientific/technical imagery that existing datasets underrepresent. The automated construction methodology also provides a template for similar dataset generation efforts.