UltraFlux：面向多宽高比高质量原生 4K 文本到图像生成的数据-模型协同设计
UltraFlux: Data-Model Co-Design for High-quality Native 4K Text-to-Image Generation across Diverse Aspect Ratios
针对原生 4K 文生图，提出 Resonance 2D RoPE 与 YaRN 结合的位置编码、非对抗 VAE 后训练、SNR 感知的 Huber 小波损失以及分阶段美学课程学习，系统解决高分辨率下位置编码、VAE 压缩与优化耦合失效的问题，可直接迁移到其他高分辨率扩散模型。
原文依据 we show that extending them to native 4K across diverse aspect ratios exposes a tightly coupled failure mode spanning positional encoding, VAE compression, and optimization
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论扩散 Transformer 近期在约 1K 分辨率下实现了强大的文本到图像生成，但我们表明，将其扩展到跨多种宽高比的原生 4K 会暴露出一个紧密耦合的失效模式，涉及位置编码、变分自编码器压缩和优化。孤立地处理其中任何一个因素都会在质量上留下大量提升空间。因此，我们采取数据-模型协同设计的视角，提出 UltraFlux，一个基于 Flux 的 DiT，在 MultiAspect-4K-1M 上原生训练于 4K，该数据集包含 100 万张 4K 图像，具有受控的多宽高比覆盖、双语描述以及丰富的 VLM/IQA 元数据，用于分辨率和宽高比感知的采样。在模型方面，UltraFlux 结合了：(i) 带有 YaRN 的 Resonance 2D RoPE，用于在 4K 下实现训练窗口、频率和宽高比感知的位置编码；(ii) 一种简单的非对抗变分自编码器后训练方案，提升 4K 重建保真度；(iii) 一种 SNR 感知的 Huber 小波目标，重新平衡跨时间步和频带的梯度；以及 (iv) 一种分阶段美学课程学习策略，将高美学监督集中在由模型先验主导的高噪声步上。这些组件共同产生了一个稳定、细节保持的 4K DiT，可泛化到宽、方、高宽高比。在 Aesthetic-Eval@4096 基准和多宽高比 4K 设置上，UltraFlux 在保真度、美学和对齐指标上持续优于强大的开源基线，并且——借助 LLM 提示词精炼器——匹配或超越专有的 Seedream 4.0。
英文原文摘要 1521 字符
Diffusion transformers have recently delivered strong text-to-image generation around 1K resolution, but we show that extending them to native 4K across diverse aspect ratios exposes a tightly coupled failure mode spanning positional encoding, VAE compression, and optimization. Tackling any of these factors in isolation leaves substantial quality on the table. We therefore take a data--model co-design view and introduce UltraFlux, a Flux-based DiT trained natively at 4K on MultiAspect-4K-1M, a 1M-image 4K corpus with controlled multi-AR coverage, bilingual captions, and rich VLM/IQA metadata for resolution- and AR-aware sampling. On the model side, UltraFlux couples (i) Resonance 2D RoPE with YaRN for training-window-, frequency-, and AR-aware positional encoding at 4K; (ii) a simple, non-adversarial VAE post-training scheme that improves 4K reconstruction fidelity; (iii) an SNR-Aware Huber Wavelet objective that rebalances gradients across timesteps and frequency bands; and (iv) a Stage-wise Aesthetic Curriculum Learning strategy that concentrates high-aesthetic supervision on high-noise steps governed by the model prior. Together, these components yield a stable, detail-preserving 4K DiT that generalizes across wide, square, and tall ARs. On the Aesthetic-Eval@4096 benchmark and multi-AR 4K settings, UltraFlux consistently outperforms strong open-source baselines across fidelity, aesthetic, and alignment metrics, and--with a LLM prompt refiner--matches or surpasses the proprietary Seedream 4.0.