UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (2510.20661v1)

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}.

• The paper introduces UltraHR-100K as a rigorously curated dataset of 100K ultra-high-resolution images with rich captions for text-to-image synthesis.
• The paper proposes frequency-aware post-training methods, DOTS and SWFR, which optimize high-frequency detail generation in diffusion models.
• The paper demonstrates superior quantitative and qualitative performance, setting new benchmarks in both local detail fidelity and overall semantic accuracy.

UltraHR-100K: A Large-Scale High-Quality Dataset and Frequency-Aware Training for Ultra-High-Resolution Image Synthesis

Introduction

Ultra-high-resolution (UHR) text-to-image (T2I) synthesis remains a challenging frontier due to the scarcity of large, high-quality datasets and the lack of training strategies specifically tailored for fine-grained detail generation. The "UltraHR-100K" paper addresses these limitations by introducing a rigorously curated dataset of 100,000 UHR images with rich captions and proposing a frequency-aware post-training method to enhance the detail synthesis capabilities of diffusion-based T2I models. The work demonstrates significant improvements in both global semantic fidelity and local detail quality, establishing new benchmarks for UHR image generation.

Figure 1: UltraHR-100K (left) is a large-scale, high-quality dataset for UHR image synthesis, enabling the generation of high-fidelity UHR images (right).

UltraHR-100K Dataset Construction

The UltraHR-100K dataset is constructed through a multi-stage pipeline designed to ensure not only high resolution but also content diversity, detail richness, and aesthetic quality. The initial collection phase gathers approximately 400K high-resolution images (minimum 3840×2160) from web sources and imaging devices. Subsequent filtering stages employ:

• Laplacian variance and Sobel operator for sharpness and edge density,
• GLCM metrics for texture complexity,
• Shannon entropy for content diversity,
• LAION Aesthetic Predictor for perceptual quality.

The final dataset is the intersection of the top 50% of images in each criterion, ensuring that every image meets stringent standards across all dimensions.

Figure 2: Left: All images in UltraHR-100K have a minimum resolution of 3K. Middle: The dataset covers a broad range of categories. Right: Caption lengths are significantly longer than Aesthetic-4K, providing richer semantic supervision.

Compared to Aesthetic-4K, UltraHR-100K is an order of magnitude larger and features more diverse and semantically rich captions, generated using Gemini 2.0, a state-of-the-art vision-LLM. This results in improved semantic alignment and expressiveness for downstream generative models.

Figure 3: UltraHR-100K provides more expressive and detailed captions than Aesthetic-4K, enhancing semantic alignment.

Frequency-Aware Post-Training: DOTS and SWFR

The paper introduces a frequency-aware post-training (FAPT) strategy to address the inability of pre-trained T2I models to synthesize fine-grained details at UHR. FAPT consists of two key components:

Detail-Oriented Timestep Sampling (DOTS)

DOTS leverages the observation that high-frequency details are synthesized in the later stages of the diffusion denoising process. By sampling denoising timesteps from a Beta($\alpha$, $\beta$) distribution (with $\alpha=2$, $\beta=4$), the training process is biased toward later steps, focusing learning on detail-critical phases. Ablation studies confirm that this configuration optimally balances detail learning and semantic consistency.

Soft-Weighting Frequency Regularization (SWFR)

SWFR introduces a frequency-domain loss by applying a 2D Discrete Fourier Transform (DFT) to both model predictions and targets. A soft weighting function $w(\mathbf{r})$ emphasizes high-frequency components, encouraging the model to reconstruct fine details without sacrificing low-frequency structure. The total loss combines the standard diffusion loss with the frequency-domain regularization, controlled by a balancing coefficient.

This approach contrasts with prior DWT-based methods (e.g., Diffusion4K), offering finer, globally coherent frequency supervision and improved high-frequency fidelity.

Experimental Results

Quantitative and Qualitative Evaluation

Experiments are conducted on the UltraHR-eval4K benchmark (4096×4096), with comparisons against SOTA training-free and training-based methods, including FLUX+BSRGAN, Pixart-$\sigma$, SANA, and Diffusion4K. Metrics include FID, FID-patch, IS, IS-patch, CLIP, and FG-CLIP.

• The proposed method achieves the best FID-patch (15.795) and competitive FID (31.748), indicating superior local detail and overall image quality.
• CLIP and FG-CLIP scores remain on par with the best baselines, demonstrating robust text-image alignment.

  Figure 4: Qualitative comparisons on UltraHR-eval4K (4096×4096) show that the proposed method generates visually complex images with rich semantic content.

  

  Figure 5: The method produces realistic textures and fine-grained details, outperforming SOTA baselines in local fidelity.

  

  Figure 6: Additional visual comparisons confirm consistent high-quality results across diverse prompts and categories.

A user paper further corroborates these findings, with the proposed method preferred in 70% of cases for overall quality, detail, and alignment.

Ablation and Analysis

Ablation studies validate the contributions of DOTS and SWFR. Removing either component degrades FID and FID-patch, with the full dataset and both modules yielding the best results. Data scale is also critical: reducing the dataset to 15K images results in a significant performance drop, underscoring the necessity of large-scale, high-quality UHR data.

Parameter sweeps for DOTS confirm that $(\alpha=2, \beta=4)$ provides the optimal trade-off between detail and semantic structure.

Implications and Future Directions

UltraHR-100K establishes a new standard for UHR T2I datasets, enabling the training and evaluation of models at unprecedented scale and quality. The frequency-aware post-training strategy demonstrates that targeted supervision in both the temporal (denoising steps) and frequency domains is essential for high-fidelity UHR synthesis.

Practically, these advances facilitate applications in digital art, virtual content creation, and commercial design, where fine detail and semantic accuracy are paramount. Theoretically, the work highlights the importance of frequency-domain regularization and adaptive training schedules for scaling generative models to higher resolutions.

Future work should address the slight degradation in text-image alignment observed with aggressive frequency regularization and expand the dataset to include more portrait data, further improving generalization and performance in underrepresented categories.

Conclusion

UltraHR-100K provides a rigorously curated, large-scale dataset and a principled frequency-aware post-training strategy, together setting new benchmarks for ultra-high-resolution text-to-image synthesis. The combination of data quality, semantic richness, and targeted training yields substantial improvements in both global and local image fidelity, with broad implications for the development and deployment of next-generation generative models.