PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Chen Zhao, En Ci, Yunzhe Xu, Tiehan Fan, Shanyan Guan, Yanhao Ge, Jian Yang, Ying Tai

arXiv:2510.20661v1cs.CV

TL;DR

UHR text-to-image generation faces a shortage of large, high-quality datasets and tailored strategies for synthesizing fine detail. The paper builds UltraHR-100K and proposes frequency-aware post-training; experiments report improved fine-detail accuracy and visual fidelity, while noting alignment and portrait-data limitations.

  • Problem

    UHR T2I research lacks a large-scale, high-quality dataset and tailored training strategies for fine-grained detail synthesis.

    Source: §1 · p. 2

  • Method

    The paper constructs UltraHR-100K and proposes frequency-aware post-training with DOTS and SWFR for pretrained T2I models.

    Source: §1 · pp. 2–3

  • Results

    Experiments on UltraHR-eval4K report improved UHR visual fidelity and fine-detail accuracy, with stronger perceptual results and competitive CLIP scores.

    Source: §5.2 · p. 9, §6 · p. 10

  • Takeaways & Limitations

    The authors conclude that their dataset and post-training approach boost the visual fidelity and fine-detail accuracy of UHR image synthesis.

    Source: §6 · p. 10

    The post-training strategy slightly degrades text–image alignment, and limited portrait data constrains improvement in UHR portrait generation.

    Source: §6 · p. 10

Abstract

from arXiv · show

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}.

1 Introduction

UHR text-to-image generation remains constrained by limited high-quality training data and insufficient attention to fine-detail synthesis. The paper introduces a curated 100K-image dataset and a frequency-aware post-training strategy, then evaluates the approach on UltraHR-eval4K.

Motivation

  • Existing T2I models often degrade or develop structural artifacts when scaled beyond their typical 1024×1024 resolution, limiting applications that require fine detail and visual fidelity.The introduction also identifies limited high-quality UHR data and a lack of detail-oriented training strategies as challenges for training-based methods.Source: §1 · p. 2

Dataset contribution

  • UltraHR-100K contains 100K UHR images selected for detail richness, content complexity, and aesthetic quality, with rich captions and a minimum resolution above 3K.The dataset is approximately 10× larger than Aesthetic-4K; the captioning process uses Gemini 2.0.Source: §1 · p. 2, Figure 2 · p. 2, Figure 3 · p. 3

Training strategy

  • The proposed frequency-aware post-training method combines DOTS, which focuses learning on detail-critical denoising steps, with SWFR, which softly constrains frequency bands using DFT.The paper contrasts SWFR's continuous frequency spectrum with Diffusion4K's discrete, block-based DWT decomposition.Source: §1 · p. 3

Evaluation

  • Experiments on UltraHR-eval4K demonstrate the method's effectiveness, with the paper reporting improved fine-grained detail representation in UHR image generation.The benchmark is 4096 × 4096; the introduction does not provide quantitative outcome values.Source: §1 · p. 3

2 Related Work

The related work traces progress in diffusion-based T2I generation and reviews approaches to UHR synthesis. It emphasizes the difficulty of scaling to 4K while preserving fine detail.

2.1 Text-to-Image Synthesis

  • Diffusion-based T2I methods progressed from iterative pixel-space denoising to latent diffusion and transformer-based models, improving scalability and generation quality.The section names DDPM and DDIM as early methods and latent diffusion as a later shift.Source: §2.1 · pp. 3–4

2.2 Ultra-High-Resolution Image Synthesis

  • Current latent diffusion models typically operate at up to 1024 × 1024, while 4K generation raises computational demands because cost grows quadratically with image size.The section describes training-free methods that modify inference and training-based models designed for more efficient 4K generation.Source: §2.2 · p. 4
  • Training-free UHR approaches can produce oversmoothed or implausible details and require prolonged inference, while existing models are also described as overlooking high-fidelity detail synthesis.Examples include upscaling and re-denoising, cascaded generation, and efficient token compression.Source: §2.2 · p. 4

3 Constructing UltraHR-100K

UltraHR-100K is curated from high-resolution images using complementary filters for texture detail, content complexity, and aesthetic quality. Its captions combine global descriptions with fine-grained details.

Data collection

  • The dataset is curated from approximately 400K collected high-resolution images, applying quality criteria beyond resolution to produce a 100K-level UHR T2I dataset.The source images have a minimum resolution of 3840×2160.Source: §3 · p. 4

Detail Richness

  • Detail-richness filtering retains the top 50% of the preliminarily cleaned images by aggregated GLCM scores for contrast, entropy, and correlation.The metrics assess spatial pixel relationships associated with texture complexity.Source: §3 · p. 4

Content Complexity

  • Content-complexity filtering retains the top 50% of cleaned images by Shannon entropy, used as a proxy for varied pixel intensities.Source: §3 · p. 4

Aesthetic Quality

  • Aesthetic filtering ranks cleaned images with the LAION Aesthetic Predictor and retains the top 50% by its perceptual-quality score.The score reflects visual composition, color harmony, and overall appeal.Source: §3 · p. 5

UltraHR-100K

  • The final dataset is the intersection of the detail, complexity, and aesthetic subsets, so each image meets all three selection criteria.The paper also constructs UltraHR-eval4K from 2,000 dataset images and reports comparisons with Aesthetic-4K and PixArt-30K.Source: §3 · p. 5

UHR Image Caption

  • Gemini 2.0 generates captions that combine global image summaries with fine-grained descriptions, providing more semantic detail than captions in Aesthetic-4K.The captioning approach addresses the limited semantic expressiveness attributed to short captions in existing datasets.Source: §3 · p. 5, Figure 4 · p. 5

4 Frequency-Aware Post-Training

FAPT targets pretrained T2I models’ difficulty synthesizing fine-grained details at UHR by combining detail-oriented timestep sampling (DOTS) with soft-weighting frequency regularization (SWFR).

4.1 Detail-Oriented Timestep Sampling

  • DOTS samples timesteps from an adjustable Beta distribution, emphasizing later denoising steps where fine-grained details are synthesized.The parameters α and β control the sampling bias.Source: §4.1 · p. 6

4.2 Soft-Weighting Frequency Regularization

  • SWFR applies DFT-based frequency regularization to prediction and target, softly weighting frequency consistency to emphasize high-frequency details alongside diffusion loss.The weighting function boosts high-frequency supervision.Source: §4.2 · pp. 6–7
  • On UltraHR-eval4K at 4096×4096, qualitative comparisons report visually complex images with rich semantic content versus previous methods.Source: Figure 6 · p. 7

5 Experiments

On UltraHR-eval4K, the method outperforms baselines on key perceptual metrics and produces stronger global semantics and local detail. Ablations support the contributions of DOTS, SWFR, and the full-scale dataset; evaluations also include a public benchmark and user study.

5.1 Implementation Details

  • Training uses two stages: 4K iterations of UltraHR-100K fine-tuning for semantic planning, then 8K iterations of frequency-aware post-training for high-frequency detail learning.Experiments use SANA on four H20 GPUs; comparisons include super-resolution, training-free, and training-based baselines, evaluated with text-alignment, overall-quality, and patch-level metrics.Source: §5.1 · pp. 7–8

5.2 Comparison to State-of-the-Art Methods

  • The method achieves superior FID, FID-patch, and IS scores, competitive CLIP scores, and a substantial FID-patch improvement on UltraHR-eval4K.Table 3 reports the quantitative comparison; the paper highlights the FID-patch gain as evidence of improved fine-detail synthesis.Source: Table 3 · p. 9, §5.2 · p. 9
  • On UltraHR-eval4K, five volunteers rated 50 randomly selected cases, and the method was superior in overall quality, detail quality, text-image alignment, and preference.Table 4 reports the user-study results.Source: Table 4 · p. 9, §5.2 · p. 9
  • Qualitative comparisons show stronger semantic richness and spatial layout, plus sharper structures and more realistic textures that preserve high-frequency information.Figure 7 compares fine-grained textures and local details on 4096 × 4096 images.Source: Figure 7 · p. 8, §5.2 · p. 9
  • On Aesthetic-Eval@4096, the method showed superior performance on 195 image-text pairs, although the limited sample count made reported FID scores relatively high.The subset contains images with a short side greater than 4096 pixels.Source: §5.2 · p. 9, Table 4 · p. 9

5.3 Ablation Study

  • Adding DOTS to the baseline improves FID and patch-level FID, while adding SWFR further improves results, particularly patch-level FID; the full dataset also outperforms a 15K subset.The C-versus-D comparison uses the same training strategy and tests the impact of data scale.Source: Table 5 · p. 10, §5.3 · p. 10
  • DOTS samples later denoising steps with α = 2 and β = 4 to emphasize detail refinement; other settings can weaken detail learning, semantic consistency, or diversity.Table 6 analyzes DOTS; the paper describes α = 2, β = 4 as a balanced configuration.Source: §5.3 · p. 10

6 Conclusion

The paper introduces a curated UHR dataset and a frequency-aware post-training method, with experiments reporting improved visual fidelity and fine-detail accuracy. It also identifies text-image alignment and limited portrait data as limitations.

6 Conclusion

  • UltraHR-100K contains 100K curated UHR images with rich textual annotations, selected for detail, visual complexity, and aesthetic appeal.The accompanying frequency-aware post-training method combines DOTS and SWFR.Source: §6 · p. 10
  • Experiments on UltraHR-eval4K confirm improved visual fidelity and fine-detail accuracy in UHR image synthesis.Source: §6 · p. 10
  • The method slightly degrades text-image alignment, and limited portrait data constrains improvement in UHR portrait generation.The authors plan to develop more balanced training strategies and add high-quality UHR portrait images.Source: §6 · p. 10
Loading 2510.20661v1…