In this paper, we present a pipeline for automatically generating a large-scale, high-quality image editing dataset to address the limitations of generative model-based image editing assistants that perform image editing via natural language commands. While existing approaches struggle to obtain accurate pixel-level editing examples, our pipeline automatically generates high-quality triplet data (original images, commands, and edited images) by directly evaluating command compliance and aesthetic factors using publicly available generative models and the Gemini validator. We increase the dataset size by 2.2 times using inversion and compositional bootstrapping techniques, and present the NHR-Edit dataset consisting of 358,000 high-quality triplets and a fine-tuned Bagel-NHR-Edit model based on it. Large-scale cross-dataset evaluations show that the proposed dataset and model outperform other publicly available datasets and models.
