UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset

TL;DR

UltraHR-100K addresses two key obstacles in ultra-high-resolution T2I: dataset scarcity and poor fine-grained detail synthesis. The authors release UltraHR-100K and propose FAPT, combining DOTS and SWFR to improve high-frequency detail in diffusion-based synthesis. Experiments on UltraHR-eval4K show improvements in texture, fidelity, and semantic alignment, validating the approach. The work provides a scalable resource and practical training techniques to enable high-quality UHR image generation for real-world applications.

Abstract

Ultra-high-resolution (UHR) text-to-image (T2I) generation has seen notable progress. However, two key challenges remain : 1) the absence of a large-scale high-quality UHR T2I dataset, and (2) the neglect of tailored training strategies for fine-grained detail synthesis in UHR scenarios. To tackle the first challenge, we introduce \textbf{UltraHR-100K}, a high-quality dataset of 100K UHR images with rich captions, offering diverse content and strong visual fidelity. Each image exceeds 3K resolution and is rigorously curated based on detail richness, content complexity, and aesthetic quality. To tackle the second challenge, we propose a frequency-aware post-training method that enhances fine-detail generation in T2I diffusion models. Specifically, we design (i) \textit{Detail-Oriented Timestep Sampling (DOTS)} to focus learning on detail-critical denoising steps, and (ii) \textit{Soft-Weighting Frequency Regularization (SWFR)}, which leverages Discrete Fourier Transform (DFT) to softly constrain frequency components, encouraging high-frequency detail preservation. Extensive experiments on our proposed UltraHR-eval4K benchmarks demonstrate that our approach significantly improves the fine-grained detail quality and overall fidelity of UHR image generation. The code is available at \href{https://github.com/NJU-PCALab/UltraHR-100k}{here}.

Figures (8)

• Figure 1: Our UltraHR-100K (left) is a large-scale high-quality dataset for ultra-high-resolution (UHR) image synthesis, featuring a diverse range of categories. Utilizing this dataset enables the generation of high-fidelity UHR images (right).
• Figure 2: We perform a rigorous selection of UltraHR-100K by evaluating all collected images across three key dimensions: detail richness, content complexity, and aesthetic quality. Left: We present representative low-quality (bad case) examples for each dimension along with their corresponding scores, highlighting the necessity of such filtering. Right: In contrast, our UltraHR-100K exhibit superior texture details, semantic complexity, and aesthetic appeal.
• Figure 3: Left: Resolution distribution of our UltraHR-100K. All images have a minimum resolution of 3K, defined as the average of height and width exceeding 3000 pixels. Middle: Image categories across our dataset. The proportion of each category mirrors its distribution in our dataset.Right: Caption length distribution. Compared to the recent Aesthetic-4Kzhang2025diffusion, our captions are significantly longer, providing richer semantic supervision.
• Figure 4: Comparison between our UltraHR-100K and Aesthetic-4Kzhang2025diffusion. Captions in our UltraHR-100K provide more expressive descriptions, encompassing not only global summaries of the image content but also rich details that enhance semantic alignment.
• Figure 5: Relation between weighting ratio and timesteps with beta sampling strategy.