Source-linked AI summary

Diffusion-4K: Ultra-High-Resolution Image Synthesis with Latent Diffusion Models

Jinjin Zhang, Qiuyu Huang, Junjie Liu, Xiefan Guo, Di Huang

arXiv:2503.18352v2cs.CV

TL;DR

Direct 4K image synthesis lacks comprehensive datasets and fine-detail evaluation while remaining computationally challenging. Diffusion-4K introduces the Aesthetic-4K benchmark and wavelet-based fine-tuning for latent diffusion models, with experiments reporting effective photorealistic 4K generation and prompt adherence, especially with SD3 and Flux.

  • Problem

    Direct training and assessment of 4K images remain limited, while conventional metrics inadequately evaluate fine details in ultra-high-resolution synthesis.

  • Method

    Diffusion-4K combines the Aesthetic-4K dataset and benchmark metrics with wavelet-based fine-tuning that emphasizes high-frequency details while preserving low-frequency approximations.

  • Results

    Diffusion-4K demonstrates effective photorealistic 4K image synthesis and superior performance, with advantages in fine-detail metrics and preference evaluations.

  • Takeaways & Limitations

    The framework supports 4K training and generation across latent diffusion models, including SD3 and Flux, while enabling comprehensive assessment of image quality and detail.

Abstract

from arXiv · show

In this paper, we present Diffusion-4K, a novel framework for direct ultra-high-resolution image synthesis using text-to-image diffusion models. The core advancements include: (1) Aesthetic-4K Benchmark: addressing the absence of a publicly available 4K image synthesis dataset, we construct Aesthetic-4K, a comprehensive benchmark for ultra-high-resolution image generation. We curated a high-quality 4K dataset with carefully selected images and captions generated by GPT-4o. Additionally, we introduce GLCM Score and Compression Ratio metrics to evaluate fine details, combined with holistic measures such as FID, Aesthetics and CLIPScore for a comprehensive assessment of ultra-high-resolution images. (2) Wavelet-based Fine-tuning: we propose a wavelet-based fine-tuning approach for direct training with photorealistic 4K images, applicable to various latent diffusion models, demonstrating its effectiveness in synthesizing highly detailed 4K images. Consequently, Diffusion-4K achieves impressive performance in high-quality image synthesis and text prompt adherence, especially when powered by modern large-scale diffusion models (e.g., SD3-2B and Flux-12B). Extensive experimental results from our benchmark demonstrate the superiority of Diffusion-4K in ultra-high-resolution image synthesis.

1. Introduction

Diffusion-4K addresses the limited exploration of direct 4K synthesis by introducing a benchmark and wavelet-based fine-tuning framework for detailed, photorealistic generation.

  • Most latent diffusion models train and generate at 1024×1024, while direct 4K synthesis remains computationally demanding and underexplored.Existing 4K approaches emphasize efficiency but neglect high-frequency details and rich textures.
  • Aesthetic-4K provides a curated ultra-high-resolution dataset with GPT-4o-generated captions and metrics for both holistic quality and fine details.The benchmark combines FID, Aesthetics and CLIPScore with GLCM Score and Compression Ratio.
  • Wavelet-based fine-tuning emphasizes high-frequency components while preserving low-frequency approximations for 4K image synthesis.The method is designed for direct training with photorealistic 4K images and is applicable to latent diffusion models.
  • The method supports multiple latent diffusion models, including SD3-2B and Flux-12B, at 4096 × 4096 resolution.Experiments use open-sourced models capable of training and generating photorealistic images at 4096 × 4096.
  • Extensive experiments report effective and generalized 4K synthesis, particularly with large-scale diffusion transformers such as SD3 and Flux.The reported effectiveness concerns ultra-high-resolution image synthesis using the proposed benchmark and method.

2. Related work

Related work develops latent diffusion and transformer-based text-to-image systems, but most models remain centered on 1024 × 1024 generation. Recent approaches pursue direct 4K synthesis mainly through efficiency improvements and specialized compression or fusion strategies.

  • Latent diffusion architectures: Latent diffusion models perform diffusion in compressed VAE latent space, and DiT applies transformer architectures within this setting.The related-work passages identify transformer scalability through models including PixArt, SD3, Flux and Playground.
  • Text-to-image conditioning: Text encoders support prompt coherence, with recent models such as SD3 and Flux combining CLIP and T5-XXL for text understanding.Earlier systems use CLIP or T5-XXL separately, while newer approaches combine both.
  • Ultra-high-resolution synthesis: Advanced latent diffusion models typically train and synthesize at 1024 × 1024 because higher resolution introduces quadratic computational costs.Training-free fusion approaches have also been developed for 4K generation from existing latent diffusion models.
  • Ultra-high-resolution synthesis: PixArt-Σ generates images close to 4K using efficient token compression, while Sana supports 1024 × 1024 through 4096 × 4096 synthesis with a deep-compression VAE.These approaches target more efficient direct ultra-high-resolution generation and training.

3. Methods

Diffusion-4K combines the Aesthetic-4K benchmark with wavelet-based fine-tuning for direct photorealistic 4K synthesis. The benchmark evaluates both holistic quality and fine details, while the method decomposes latent features into frequency components for training.

  • Aesthetic-4K Benchmark: Aesthetic-4K provides a benchmark for direct ultra-high-resolution synthesis using curated images, GPT-4o captions, and qualitative and quantitative evaluation.Its evaluation combines holistic metrics with measures designed for fine details in 4K images.
  • Aesthetic-4K Benchmark: The dataset contains 12,015 curated training images and 2,781 evaluation images, with an Aesthetic-Eval@4096 subset of 195 images.The training images have median dimensions of 4128×4640, while the evaluation images have median dimensions of 2983×3613.
  • Aesthetic-4K Benchmark: GLCM Score and Compression Ratio complement FID, Aesthetics, and CLIPScore by assessing local texture richness and image detail.GLCM Score averages entropy-based measurements over 64-sized local patches, while Compression Ratio uses JPEG compression at quality 95.
  • Wavelet-based Fine-tuning: Wavelet-based fine-tuning enables direct training with photorealistic 4096×4096 images across latent diffusion models.The approach uses a partitioned VAE with down-sampling factor F = 16 to reduce memory consumption and avoid out-of-memory issues.
  • Wavelet-based Fine-tuning: The training objective applies a discrete wavelet transform to predicted and target quantities, jointly optimizing low-frequency structure and high-frequency details.Haar wavelets decompose features into LL, LH, HL, and HH sub-bands, and the reconstruction objective can be substituted to support different diffusion models.

4. Experiments

Experiments evaluate Diffusion-4K with partitioned VAEs, WLF fine-tuning, standard and fine-detail metrics, qualitative comparisons, preference studies, and ablations. Results support 4K reconstruction, generation, prompt adherence, aesthetics, and fine-detail quality.

  • Experimental setup: SD3-2B and Flux-12B are evaluated with F = 16 partitioned VAEs, standard metrics, and proposed GLCM Score and Compression Ratio metrics.The evaluation covers reconstruction, image quality, prompt adherence, and fine details at high resolution.
  • Partitioned VAE: The partitioned VAE resolves the original VAE's out-of-memory issue without retraining or fine-tuning, preserving latent-space consistency for subsequent 4K fine-tuning.Qualitative results demonstrate effective 4K image reconstruction.
  • Qualitative synthesis: Diffusion-4K synthesizes qualitative 4K images at varied aspect ratios and supports direct photorealistic generation beyond the 4096 × 4096 fine-tuning resolution.The reported qualitative examples use Flux-12B.
  • Preference study: WLF achieves a superior win rate in both human and GPT-4o preference evaluations for visual aesthetics, prompt adherence, and fine details.The preference study compares Flux outputs with and without WLF.
  • Caption ablation: GPT-4o-generated captions improve Aesthetic-Eval@4096 results for both SD3 and Flux compared with original LAION-5B captions.The ablation links higher-quality captions with improved image synthesis and prompt coherence.
  • Ablation studies: WLF outperforms fine-tuning without WLF on CLIPScore, Aesthetics, GLCM Score, and Compression Ratio.The ablation attributes gains across visual aesthetics, prompt adherence, and high-frequency details to WLF.

5. Conclusion

The conclusion presents Diffusion-4K as a framework combining the Aesthetic-4K benchmark with wavelet-based fine-tuning for ultra-high-resolution text-to-image generation. Qualitative and quantitative results support photorealistic 4K training and generation with models such as SD3 and Flux.

  • Conclusion: Diffusion-4K combines the Aesthetic-4K benchmark and wavelet-based fine-tuning for ultra-high-resolution text-to-image generation.The framework is designed for training state-of-the-art latent diffusion models such as SD3 and Flux.
  • Conclusion: Qualitative and quantitative results demonstrate effective training and generation of photorealistic 4K images.

PixArt-Sigma

The figure compares PixArt-Σ and Diffusion-4K using identical prompts, placing PixArt-Σ images on the left and Diffusion-4K images on the right. The highlighted patches emphasize fine-detail differences.

  • Comparison layout: Diffusion-4K images appear on the right and PixArt-Σ images on the left under identical prompts.
  • Fine-detail comparison: Yellow and red patches identify corresponding fine-detail regions in the comparison.The passage states that the comparison demonstrates Diffusion-4K's superiority in fine details.

7. Comparisons

Diffusion-4K is compared with direct ultra-high-resolution methods and shown qualitatively across diverse prompts, aspect ratios, seeds, and text-rendering cases. Wavelet-based latent fine-tuning preserves structure while enriching fine details.

  • Performance Comparisons: Diffusion-4K consistently produces impressive fine details under identical prompts, highlighting superiority over PixArt-Σ in 4K generation.PixArt-Σ generates only 3840 × 2160 images, while the comparison uses images from PixArt-Σ and Sana's official websites.
  • Qualitative Evaluation of WLF: Wavelet-based latent fine-tuning decomposes latents into high- and low-frequency components to refine details while preserving overall structure.The decomposition is presented as making fine-tuning efficient and precise.
  • Qualitative Evaluation of WLF: Images generated with WLF exhibit richer details than images generated without WLF.
  • More Results of 4K Image Synthesis: Generated 4K images demonstrate visual aesthetics, text-prompt adherence, and fine details across different text prompts.
  • More Results of 4K Image Synthesis: Varying aspect ratios, random seeds, and spelled texts illustrate diversity in the synthesized 4K images.

8. Aesthetic-4K Dataset

Aesthetic-4K is presented as a high-quality, diverse 4K image-text dataset whose resolution substantially exceeds earlier open-source data. Its samples and statistics document both visual richness and textual coverage.

  • Dataset Statistics: Aesthetic-Train has median height and width of 4128 and 4640 pixels, compared with 1615 and 1801 pixels for PixArt-30k.The LAION-Aesthetics dataset contains approximately 0.03% 4K images, underscoring the scarcity of such data in open-source datasets.
  • Dataset Statistics: Aesthetic-4K provides histograms of image height and width to characterize its resolution distribution.
  • Dataset Statistics: Caption word clouds visualize the textual distribution of the Aesthetic-4K dataset.
  • Qualitative Samples: Training samples pair ultra-high-resolution images with precisely generated GPT-4o captions, illustrating dataset diversity and richness.

9. More Details

The paper details 4K preprocessing, latent fine-tuning, scalable-model training, and evaluation procedures. Additional figures and tables document captioning, dataset statistics, reconstruction, efficiency, and human assessment.

  • Training Details: Images are resized so the shorter dimension is 4096, randomly cropped to 4096 × 4096, and normalized using mean and standard deviation 0.5.
  • Training Details: The partitioned VAE compresses pixel space into a latent space with F = 16.
  • Training Details: Latent diffusion models are optimized with the wavelet-based latent fine-tuning objective while text encoders and the partitioned VAE remain fixed.Text embeddings are pre-computed, and the default DiT patch size is P = 2 for SD3-2B and Flux-12B.
  • Training Efficiency: Fine-tuning a 12B diffusion model requires approximately 2,000 A100 GPU hours.The training details cover SD3-2B and Flux-12B at 4096 × 4096 resolution.
  • Reconstruction: Reconstruction results at 512 × 512 showcase the partitioned VAE's image-handling capability.
  • Evaluation Details: FID, Aesthetics, and CLIPScore are computed using images resized to a fixed low resolution.
  • Human Evaluation: Five participants rate extracted patches from 1 to 10 for visual details, with averages used to calculate SRCC and PLCC.Compression Ratio correlations use its reciprocal.
  • Human Evaluation: Human preference evaluation uses 112 prompts and ten participants rating visual aesthetics, prompt adherence, and fine details.
Loading 2503.18352v2…