UltraHR-100K, released in 2025 by Nanjing University in collaboration with vivo, is a large-scale, high-quality dataset for ultra-high resolution (UHR) text-to-image (T2I) tasks. The related paper, "UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset," has been accepted by NeurIPS 2025 and aims to improve the ability of diffusion models in fine-grained detail synthesis, content diversity representation, and visual fidelity.

This dataset contains approximately 100,000 ultra-high-resolution images, covering a wide range of themes including people, architecture, natural landscapes, art illustrations, and product design. Each image has a resolution exceeding 3K and is accompanied by high-quality rich text descriptions. The data is sourced from various publicly available resources and underwent a rigorous three-stage quality control process. The selection criteria included richness of detail, scene/content complexity, and aesthetic quality, ensuring that the images possess high credibility and strong expressiveness in terms of texture detail, structural hierarchy, and overall style. It can serve as a standard benchmark for training, fine-tuning, and evaluating ultra-high-resolution generative models.