UltraFlux Achieves High-Quality Native 4K Text-to-Image Generation

Generating high-resolution images with text-to-image models, especially at 4K across various aspect ratios, presents significant challenges for current diffusion transformers. Issues arise from positional encoding, VAE compression, and optimization, which are tightly coupled and cannot be solved in isolation. A new approach, UltraFlux, tackles these problems through a data-model co-design strategy. It utilizes a Flux-based Diffusion Transformer trained on a specially curated 1-million-image 4K dataset, MultiAspect-4K-1M, which includes diverse aspect ratios and rich metadata. UltraFlux incorporates several innovations: Resonance 2D RoPE for positional encoding, a non-adversarial VAE post-training scheme for better 4K reconstruction, an SNR-Aware Huber Wavelet objective for gradient rebalancing, and a Stage-wise Aesthetic Curriculum Learning strategy. These combined elements enable stable, detail-preserving 4K image generation that generalizes across wide, square, and tall aspect ratios, surpassing open-source baselines and matching proprietary models.