PaperlayerView exact revision on arXiv

Source-linked AI summary

PixVerve: Advancing Native UHR Image Generation to 100MP with a Large-Scale High-Quality Dataset

Haojun Chen, Haoyang He, Chengming Xu, Qingdong He, Junwei Zhu, Yabiao Wang, Zhucun Xue, Xianfang Zeng, Zhennan Chen, Xiaobin Hu, Hao Zhao, Yong Liu, Jiangning Zhang, Dacheng Tao

arXiv:2605.20147v1cs.CV

TL;DR

Native ultra-high-resolution text-to-image generation remains challenging because models trained at 1K–2K can suffer structural artifacts, repetition, and lost details when extrapolated. PixVerve introduces a 95K-image 100MP dataset, native-generation training schemes, and a dedicated benchmark, finding consistent gains from more descriptive captions across evaluated models.

  • Problem

    Existing T2I models focus on 1K–2K resolutions, while direct extrapolation to UHR generation can produce artifacts, repetition, and lost high-frequency detail.

  • Method

    The paper builds PixVerve-95K, explores training schemes for native 100MP generation, and introduces PixVerve-Bench for UHR evaluation.

  • Results

    Long captions consistently improve 4K generation performance across UltraFlux, Diffusion-4K, and FLUX.2-I (Full) on PixVerve-Bench.

  • Takeaways & Limitations

    The framework provides a dataset, training exploration, and UHR-specific benchmark for studying and evaluating native 100MP T2I generation.

  • Takeaways & Limitations

    PixVerve-95K remains smaller than existing general-purpose T2I datasets, and UHR-specific architectures and more efficient, robust training recipes remain underexplored.

Abstract

from arXiv · show

Text-to-Image (T2I) models have recently seen notable progress around 1K and 2K resolution. With the extreme desire for better visual experience and the rapid development of imaging technology, the demand for Ultra-High-Resolution (UHR) image generation has grown significantly. However, UHR image generation poses great challenges due to the scarcity and complexity of high-resolution content. In this paper, we first introduce PixVerve-95K, a high-quality, open-source UHR T2I dataset curated with a carefully designed data pipeline, which contains 95K images across diverse scenarios (each image has a minimum pixel-count of 100M) and seven-dimensional annotations. Based on our large-scale image-text dataset, we take a pioneering step to extend various T2I foundation models to native 100MP generation with three training schemes. Finally, leveraging both conventional metrics and multimodal large language model-based assessments, our proposed PixVerve-Bench benchmark establishes a comprehensive evaluation protocol for UHR images encompassing visual quality and semantic alignment. Extensive experimental results on our benchmark and the constructive exploration of training strategies collaboratively provide valuable insights for future breakthroughs.

1 Introduction

Existing T2I models largely target 1K–2K resolutions, while growing demand for gigapixel-scale content makes native UHR generation increasingly important and technically challenging. PixVerve addresses these gaps with a 100MP dataset, native-generation training schemes, and a comprehensive benchmark.

  • Most existing T2I models train and generate at fixed 1K or 2K resolutions, and direct extrapolation to UHR causes structural artifacts and quality degradation.
  • Growing demand for gigapixel-scale content is driven by next-generation media, imaging advances, and native 100MP smartphone specifications.
  • Training-based methods improve UHR detail synthesis by capturing long-range image correlations, but extending native generation beyond 4K remains challenging.
  • 95,735 100MP images comprise PixVerve-95K, curated through a five-stage automated pipeline with five metadata types and two comprehensive captions.
  • PixVerve extends latent and pixel diffusion foundation models to native 100MP generation through three training schemes and introduces PixVerve-Bench with hierarchical conventional and MLLM-based evaluation.

2 Related Work

Related work traces T2I progress from web-scale image-text datasets and evolving architectures toward quality-focused data and diffusion-based models. Beyond 2K, existing UHR methods either scale images without training or train for native generation, but remain below the 100MP frontier and can suffer quality limitations.

  • Datasets: Large-scale datasets such as LAION-400M and LAION-5B enabled T2I foundation-model training, while dataset construction increasingly prioritizes quality over volume.Higher-resolution and visual-fidelity demands further motivate quality-focused dataset development.
  • T2I architectures: T2I architectures evolved from GANs and autoregressive models to diffusion models, with latent diffusion reducing computation while preserving perceptual fidelity.Diffusion models became the prevailing paradigm, and latent diffusion performs diffusion in compressed latent space.
  • UHR generation: Beyond 2K, UHR generation is dominated by latent diffusion models and follows training-free scaling or training-based native-generation paradigms.Training-free methods are more resource-friendly, whereas training-based methods target improved synthesis quality.
  • UHR generation: Training-free UHR scaling often produces object repetition, texture degradation, and unrealistic details despite lower resource requirements.These limitations motivate training-based approaches for higher-quality synthesis.
  • UHR generation: Existing native-UHR solutions remain confined to sub-4K or 4K resolution, falling short of the gigapixel-scale fidelity required for real-world applications.The stated objective is to push T2I generation to the 100MP scale.

3 Methodology: Dataset, Model, and Benchmark

The methodology addresses native 100MP text-to-image generation through an integrated dataset, training, and evaluation regime. It introduces PixVerve-95K, a curated high-resolution corpus with structured annotations, and PixVerve-Bench, a holistic benchmark combining conventional metrics with MLLM-based assessment.

  • Methodology Overview: Native 100MP generation is treated as an end-to-end training and evaluation regime requiring intrinsic high-resolution priors, rather than training-free upscaling.The methodology targets both the scarcity of high-quality 100MP T2I data and the lack of a systematic UHR benchmark.
  • Dataset Construction: 95,735 images comprise PixVerve-95K, an open-source ∼100MP corpus with comprehensive annotations extending beyond long and short captions.The dataset provides five-dimensional metadata covering basic visual scores, tags, bounding boxes, aesthetics-level analysis, and instance-level descriptions, supporting granular quality control and adaptive sampling.
  • Benchmark: PixVerve-Bench contains 200 manually selected images averaging 12369×14377 resolution and evaluates UHR T2I models through conventional metrics and MLLM-as-a-judge protocols.Its framework assesses complementary visual-quality and semantic-alignment aspects across diverse scenarios.
  • Benchmark: The Instance-centric Compliance Score uses Qwen3.5-35B-A3B to measure semantic alignment through instance existence, appearance attributes, and spatial relations.ICS provides a fine-grained and interpretable assessment of whether generated visual elements follow textual prompts.

4 Experiments

Experiments compare PixVerve methods and baselines across 4K, 8K, and 10K, showing distinct trade-offs between adaptation quality, scalability, speed, and detail reconstruction. Scheme I excels at 4K, Scheme II improves efficiency but degrades quality, while Scheme III offers the strongest scalability and practical native 100MP inference.

  • Experimental Setup: The evaluation compares proposed variants and direct-extrapolation, training-free, and training-based baselines at 4K, 8K, and 10K.Proposed variants are FLUX.2-I (Full), FLUX.2-I (LoRA), FLUX.2-II, and L2P-III.
  • Base Model and Existing Methods: At 8K, direct FLUX.2-klein-base-4B extrapolation yields FID above 422 and CLIPScore 18.345, while DemoFusion achieves the best FIDpatch at 8K and 10K but limited semantic alignment.DemoFusion’s ICS remains below 3.7 across all reported resolutions.
  • Scheme I: Strong 4K Adaptation but Poor Scalability: Scheme I provides the best 4K balance: LoRA reduces FIDpatch from 76.794 to 40.433, while full-parameter tuning raises ICS from 5.376 to 8.533.The results indicate improved local image statistics while preserving the semantic prior.
  • Scheme II: Faster Attention with Optimization Difficulty: Scheme II delivers a 30% 4K inference speedup and requires 9,216 H20 GPU hours for three training epochs, but reaches FIDpatch 76.460 and ICS 5.340.Its quality is close to the base model and substantially below Scheme I variants.
  • Scheme III: Best Scalability with a Patch-Size Trade-off: L2P-III achieves the best FID at 4K and 8K, scoring 118.183 and 134.635, remains functional at 10K, and uses one GPU for 58s, 70s, and 88s inference.It is up to 156× faster than DemoFusion and over 33× faster than FLUX.2-I LoRA at 10K, but larger patches weaken high-resolution detail reconstruction.
  • Ablation on Image Caption Quality: Long captions consistently improve 4K generation across UltraFlux, Diffusion-4K, and FLUX.2-I (Full), underscoring the importance of descriptive granularity and semantic density.The ablation evaluates short versus long prompting on PixVerve-Bench.

5 Conclusion · Appendix

The paper concludes by presenting an integrated framework for native 100MP text-to-image generation through data, training, and evaluation, while acknowledging dataset and research limitations. The appendix documents implementation, analysis, licensing, evaluation, and additional qualitative and quantitative results.

  • 5 Conclusion: The framework addresses native 100MP T2I generation through coordinated advances in data, training, and evaluation.It targets challenges in data scarcity and semantic granularity.
  • 5 Conclusion: PixVerve-95K is a high-quality, open-source 100MP dataset curated through a five-stage automated pipeline.The pipeline is designed to ensure data excellence.
  • 5 Conclusion: PixVerve-Bench provides UHR-specific feedback for model evaluation and selection, while the work explores training schemes extending foundation models to native 100MP generation.The benchmark and training schemes are built around PixVerve-95K.
  • 5 Conclusion: The work leaves scope for UHR-specific architectures and more efficient, robust training recipes, while PixVerve-95K remains limited in corpus size relative to general-purpose T2I datasets.The authors also emphasize vigilance, regulation, and technical responses regarding societal implications.
  • Appendix: The appendix provides implementation details for flatness detection, licensing clarification for PixVerve-95K, and additional training and inference details.These materials strengthen the main manuscript and support transparency.
  • Appendix: Additional analyses examine PixVerve-95K in the frequency domain and provide more qualitative dataset samples.These sections further assess dataset quality and content.
  • Appendix: The appendix expands PixVerve-Bench documentation with evaluation procedures, scoring formulas, human alignment analysis, and prompts used for MLLM evaluation.It also reports additional quantitative and qualitative model-performance results.
  • Appendix: Benchmark results evaluate image compression using the paper’s 100MP images.This provides an additional appendix experiment on the proposed high-resolution imagery.

A Implementation Details of Flatness Detection

Flatness detection purifies the raw data by identifying overly flat images through conservative Sobel-variance analysis over 240 × 240 patches. Images are discarded when textureless patches exceed 97.5%, with thresholds selected through manual visual audits to preserve legitimate low-texture content.

  • Flatness Detection: Each image is converted to grayscale and partitioned into 240 × 240 non-overlapping patches for flatness detection.The procedure is part of preliminary data purification.
  • Flatness Detection: Sobel operators with kernel size 3 compute each patch’s gradient magnitude, Gmag = qG2x + G2y.
  • Flatness Detection: 750 is the variance threshold below which a patch is categorized as textureless.
  • Flatness Detection: 97.5% is the textureless-patch proportion above which an image is discarded.Both thresholds are conservative and were determined empirically through manual visual audits.

B Further Analysis of PixVerve-95K Dataset

PixVerve-95K is analyzed in the frequency domain using RAPS to assess high-frequency texture realism and structural consistency. Its synthetic 100MP data matches native spectral distributions without significant high-frequency attenuation, while down-sampled SR images align with their LR counterparts.

  • Frequency-domain quality analysis: RAPS analysis evaluates PixVerve-95K’s power distribution across spatial frequencies and the realism of synthesized high-frequency textures.The analysis uses the Radially Averaged Power Spectrum to characterize frequency-domain quality.
  • Frequency-domain quality analysis: The synthetic 100MP data closely matches the native power spectral distribution across the entire frequency range.The analysis reports no significant energy attenuation in the high-frequency regime, supporting micro-texture fidelity.
  • Structural consistency: The data curation pipeline preserves global structural consistency by aligning down-sampled SR images’ power spectra with their original LR counterparts.This alignment indicates adherence to the underlying distribution of the low-resolution data.

C Licensing and Dataset Release

The dataset draws substantially on permissively licensed sources, including Pexels and Unsplash, while other subsets permit non-commercial research use. It is released under CC BY-NC 4.0 with stated compliance with copyright and privacy regulations.

  • Source Licensing: Pexels and Unsplash provide broad permissions to download, use, and modify images for commercial and non-commercial purposes without financial obligation.A significant portion of the dataset was sourced from these platforms.
  • Source Licensing: Aesthetic-Train-V2 and UltraHR-100K permit use of their data for non-commercial research purposes.
  • Dataset Release: The authors state that the dataset complies with current copyright laws and privacy regulations and release it under the CC BY-NC 4.0 license.The license is intended to prevent unauthorized commercial exploitation.

D Qualitative Samples in PixVerve-95K Dataset · E More Training and Inference Details · E.1 Training Details

The supplementary sections provide qualitative examples from PixVerve-95K and document training configurations for reproducibility. They describe resolution-specific fine-tuning, computational accounting, early termination of underperforming schemes, and initialization choices for Scheme I.

  • D Qualitative Samples in PixVerve-95K Dataset: Fig. A5 and Fig. A6 present two qualitative samples from the PixVerve-95K dataset.The samples are intended to be viewed zoomed-in.
  • E More Training and Inference Details: The supplementary material includes additional training and inference details supporting the reproducibility of the main experimental results.The provided training-details passage specifically frames these computational details as reproducibility support.
  • E.1 Training Details: For each training scheme, Table A1 reports fine-tuning epochs at each resolution scale and corresponding NVIDIA H20 GPU hours.GPU hours equal the number of NVIDIA H20 GPUs multiplied by wall-clock training time.
  • E.1 Training Details: Certain training schemes stop before completing all three stages because of unsatisfactory performance at low resolutions.The passage presents this early termination as part of the reported computational details.
  • E.1 Training Details: Scheme I compares full-parameter fine-tuning with LoRA-based parameter-efficient fine-tuning.Both approaches are evaluated within the same training scheme.
  • E.1 Training Details: The full-parameter 8K model is initialized from a 4K checkpoint trained for 3 epochs and fine-tuned for 0.25 epochs at 8K resolution.This initialization and fine-tuning sequence is specified for Scheme I.
  • E.1 Training Details: LoRA models at different resolutions are independently fine-tuned from the FLUX.2-klein-base-4B base model rather than lower-resolution LoRA checkpoints.The passage explicitly contrasts independent initialization with checkpoint continuation across resolutions.

E.2 Inference Details … F.3 Instance-centric Compliance Score (ICS)

The paper details inference efficiency across three UHR generation schemes and specifies three PixVerve-Bench metrics for texture, multi-scale fidelity, and instance-centric textual compliance. These procedures combine computational measurements with MLLM-based scoring and explicit aggregation rules.

  • E.2 Inference Details: Scheme I latency increases sharply with resolution: raising resolution from 4K to 8K increases inference time from 103s to over 1,200s, while LoRA requires nearly 3,000s at 10K.Full fine-tuning and LoRA have nearly identical latency because both retain dominant full-attention computation over the high-resolution latent grid.
  • E.2 Inference Details: Scheme II reduces 4K inference time from 103s to 71s under the same 8-GPU setting through window-attention retrofitting, but still requires multi-GPU inference.This makes it an efficiency improvement rather than a complete deployment solution.
  • E.2 Inference Details: Scheme III runs on a single GPU with 58s, 70s, and 88s at 4K, 8K, and 10K, respectively, using adaptive patch sizes to maintain nearly flat latency.It is 16.3×, 90.9×, and 155.6× faster than DemoFusion at 4K, 8K, and 10K, respectively.
  • F.1 GLCM Score: The GLCM Score quantizes grayscale intensities into 64 levels, partitions images into 64 × 64 non-overlapping patches, and averages normalized GLCM entropy across patches.Higher GLCM Scores indicate richer texture and provide an objective assessment of micro-structural complexity.
  • F.2 Multi-scale Fidelity Index (MSFI): MSFI evaluates UHR fidelity across global and local scales using nine artifact-focused sub-dimensions scored by Qwen3.5-35B-A3B on a five-point scale.Global assessment uses the resized complete image, while local assessment samples ten representative 512 × 512 patches for images below 8K resolution.
  • F.2 Multi-scale Fidelity Index (MSFI): MSFI combines weighted global and local fidelity scores, with weights derived from user-study importance ratings and a formulation designed to penalize structurally incoherent images.The score ranges from 1.2 to 10, with values approaching 10 indicating superior multi-scale fidelity.

F.4 Human Alignment for Metric Validation · G More Quantitative and Qualitative Results · G.1 More Quantitative Results

The study validates MSFI and ICS against human preferences using pairwise judgments across four T2I models at 4K and 8K. The metrics’ model rankings exactly match human rankings, while additional results detail MSFI’s nine MLLM-as-a-judge sub-dimensions.

  • F.4 Human Alignment for Metric Validation: Four representative T2I models generated images at 4K and 8K for human-alignment evaluation.The study used 30 unique prompts at each resolution.
  • F.4 Human Alignment for Metric Validation: 360 pair-wise comparison sets were evaluated by 8 participants, with every pair receiving two independent annotations.Each participant evaluated 90 image pairs.
  • F.4 Human Alignment for Metric Validation: Human judgments assessed Image Fidelity and Instance-centric Semantic Alignment, corresponding respectively to MSFI and ICS.Participants received detailed definitions and illustrative examples for both dimensions.
  • F.4 Human Alignment for Metric Validation: Human preference scores assign 1.0 to the preferred model, 0.5 to indistinguishable pairs, and 0 to the nonpreferred model.Final scores divide each model’s total score by its number of comparisons.
  • F.4 Human Alignment for Metric Validation: MSFI and ICS produce model rankings that perfectly match rankings derived from human preference scores.The consistency supports their alignment with subjective judgments of T2I quality and semantic alignment.
  • G.1 More Quantitative Results: Tab. A4 reports different methods’ performance on MSFI across nine MLLM-as-a-judge sub-dimensions.The dimensions are SC-global, PI, LC, CH, NGE, GA, TF, MGC, and SC-local.

G.2 More Qualitative Results

Figure A4 presents qualitative comparisons of 4K image generations from FLUX.2-I (Full) under both short- and long-prompting settings.

  • More Qualitative Results: Fig. A4 compares 4K image generations under short and long prompting settings, with all case pairs produced by FLUX.2-I (Full).The figure provides qualitative visual comparisons across the two prompting conditions.

H Benchmarking for Image Compression at 100MP Scale · I Prompts for MLLM Evaluation

The paper evaluates 100MP image compression under a standardized lossless-testing setup and defines MLLM-based protocols for global fidelity, local fidelity, and instance-centric semantic compliance. These protocols specify dimensions, scoring rubrics, contextual constraints, and machine-readable output formats.

  • H Benchmarking for Image Compression at 100MP Scale: The compression benchmark converts 25 dataset images to PNM and evaluates 20 lossless image formats in a single-threaded environment.Testing uses a 12th Gen Intel i7-12700H CPU, 16 GB RAM, and Windows 11; compressed size and compression time are recorded.
  • I Prompts for MLLM Evaluation: The global-fidelity prompt assesses synthetic images across structural coherence, perspective integrity, lighting consistency, and color harmony.It focuses on overall composition, geometric and physical logic, macro-scale consistency, and visual fidelity.
  • I Prompts for MLLM Evaluation: Global fidelity scores must be integers from 1 to 5 for SC-global, PI, LC, and CH, with concise reasoning returned in a JSON object inside <json> tags.The output keys are “SC-global”, “PI”, “LC”, “CH”, and “reasoning”.
  • I Prompts for MLLM Evaluation: The local-fidelity prompt evaluates a local patch across noise and grain existence, generative artifacts, texture fidelity, micro-geometry coherence, and sharpness consistency.The full global image provides theme and context, while the patch is the target for scoring and incomplete-object or composition issues are ignored.
  • I Prompts for MLLM Evaluation: Local-fidelity scoring uses a strict 1-to-5 integer scale for NGE, GA, TF, MGC, and SC-local, followed by concise reasoning in JSON format.The protocol requires independent evaluation, strict artifact sensitivity, and a high standard for a score of 5.
  • I Prompts for MLLM Evaluation: Instance-centric scores are integers from 1 to 10, with reasoning and the three metric values returned in a JSON object inside <json> tags.The rubric distinguishes complete presence, attribute fidelity, and spatial-relation accuracy across ten-point score ranges.
Loading 2605.20147v1…