AI image editors are bad at following complicated instructions. A new dataset called OpenGPT-4o-Image wants to fix that by fixing the training data, not the model.
Researchers built OpenGPT-4o-Image using an automated pipeline that pairs a structured task taxonomy with GPT-4o itself to generate the examples. The result is 80,000 instruction-image pairs spanning 11 domains and 51 subtasks, including basics like style transfer alongside harder cases: rendering text inside images, illustrating chemistry diagrams, and executing several edits in a single instruction at once. That last category is the one most existing datasets skip, since it requires a model to track multiple simultaneous changes rather than one clean edit. The team then fine-tuned existing models on the dataset and measured the effect on standard benchmarks.
The gains were not trivial. UniWorld-V1 improved by up to 18% on the ImgEdit-Bench editing benchmark, and Harmon improved by up to 13% on the GenEval generation benchmark, just from better training data on the same underlying architecture. That is a useful data point in an ongoing argument in multimodal AI: whether progress comes mainly from bigger models or from more systematically constructed training sets. This result leans toward the latter, echoing how instruction-tuning datasets reshaped text-only language models a few years back.
Worth remembering that this dataset was itself generated with GPT-4o, so it inherits whatever habits and blind spots that model has. A benchmark bump is not the same as a model that reliably nails a five-step edit request from a real user.