Research map / Image generation / Text-to-image diffusion
UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds UltraHR-100K, a 100K-image >3K-resolution text-to-image dataset with long captions, and a frequency-aware post-training method (DOTS timestep sampling plus DFT-based SWFR loss) that improves fine detail in UHR SANA generation.
Problem
Ultra-high-resolution text-to-image generation lacks a large-scale high-quality UHR dataset and training strategies tailored to fine-grained detail synthesis.
Approach
Filters 400K crawled high-resolution images by intersecting top-50% sets for GLCM detail richness, Shannon entropy complexity, and LAION aesthetic score, captions them with Gemini 2.0, and post-trains diffusion models with beta-distributed timestep sampling biased toward late denoising steps plus a DFT-domain soft-weighted high-frequency loss.
Key results
- On UltraHR-eval4K, UltraHR-100K+FAPT reaches FID 31.748 and FID-patch 15.795, versus 37.070 / 38.795 for SANA and 33.171 / 32.198 for PixArt-sigma.
- In the user study (5 volunteers, 50 cases), the method is preferred in 70% of cases for overall quality, 78% for detail quality, 72% for text-image alignment.
- Ablation: adding DOTS and SWFR to the full-data baseline lowers FID-patch from 20.93 to 15.79; training on a 15K subset instead of full data raises FID-patch to 18.42.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FAPT to UltraHR-100K training | SANA fine-tuned on UltraHR-100K · UltraHR-eval4K | FID-patch: 20.932 -> 15.795 | Helped |
| Add FAPT to UltraHR-100K training | SANA fine-tuned on UltraHR-100K · UltraHR-eval4K | IS-patch: 5.020 -> 5.104 | Hurt |
| Add FAPT to UltraHR-100K training | SANA fine-tuned on UltraHR-100K · UltraHR-eval4K | CLIP: 31.85 -> 31.82 | Hurt |
| Add DOTS while omitting SWFR (ablation) | Model A full fine-tuning on UltraHR-100K · UltraHR-eval4K | FID: 33.99 -> 32.57 | Helped |
| Add DOTS while omitting SWFR (ablation) | Model A full fine-tuning on UltraHR-100K · UltraHR-eval4K | FID-patch: 20.93 -> 19.95 | Helped |
| Add SWFR to DOTS training (ablation) | Model B full fine-tuning on UltraHR-100K · UltraHR-eval4K | FID-patch: 19.95 -> 15.79 | Helped |
| Use the full UltraHR-100K dataset instead of the randomly sampled 15K subset (ablation) | DOTS and SWFR training · UltraHR-eval4K | FID-patch: 18.42 -> 15.79 | Helped |
| Use LoRA instead of full fine-tuning (ablation) | SANA trained on full UltraHR-100K · UltraHR-eval4K | FID-patch: 20.93 -> 35.02 | Hurt |
| Set DOTS α=3, β=4 instead of α=2, β=4 (ablation) | DOTS training · UltraHR-eval4K | FID-patch: 15.795 -> 22.143 | Hurt |
Surprising findings
- The frequency-aware post-training slightly degrades text-image alignment (CLIP 31.85 baseline vs 31.82 with FAPT).
- Training-free and DWT-based baselines such as Diffusion4K score worse than SANA on several metrics despite UHR training.
Limitations
- Slight degradation in text-image alignment after frequency-aware post-training.
- Dataset contains limited portrait data, constraining UHR portrait generation.
- Training only conducted on SANA due to computational constraints.
Experimental setup
Datasets: UltraHR-100K Aesthetic-4K PixArt-30k
Benchmarks: UltraHR-eval4K Aesthetic-Eval@4096
Metrics: FID FID-patch IS IS-patch CLIP score FG-CLIP score user study preference
Baselines: PixArt-sigma SANA Diffusion4K FLUX + BSRGAN SD3.5 + BSRGAN I-Max HiFlow
Builds on: SANA Diffusion4K SD3 logit-normal sampling CAME optimizer
Closest papers by meaning
- Diffusion-4K: Ultra-High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2025 · similarity 0.89)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.89)
- Ultra-Resolution Adaptation with Ease (ICML 2025 · similarity 0.88)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.87)
- UltraVideo: High-Quality UHD Video Dataset with Comprehensive Captions (NeurIPS 2025 · similarity 0.87)
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.