An open-source image model can now generate native 4K images across wide, square, and tall formats without the quality collapse that has plagued earlier attempts.
Researchers introduced UltraFlux, a diffusion transformer built on the Flux architecture and trained on a custom dataset of one million 4K images. The dataset, called MultiAspect-4K-1M, includes bilingual captions and metadata designed to help the model handle different aspect ratios without degrading. On the technical side, the team fixed four interlocking problems simultaneously: positional encoding that breaks at high resolution, a compression stage that loses fine detail, a training objective that weights gradient signals poorly, and an aesthetic curriculum that front-loads quality supervision where it counts. Their finding is that fixing any one of these in isolation leaves significant quality on the table.
The result matches or beats Seedream 4.0 — a proprietary model from ByteDance — on aesthetic and alignment benchmarks, which is a meaningful threshold: proprietary models have held a visible quality lead over open alternatives at very high resolutions. The co-design framing, treating data and model architecture as a single system rather than separate problems, could give other open-source teams a cleaner roadmap for similar resolution jumps.
The paper is an arXiv preprint, so peer review is pending — and benchmark comparisons against a competitor's model, especially one that includes a prompt-rewriting step to close the gap, deserve the usual scrutiny before anyone calls this a settled contest.