UltraImage: Rethinking Resolution Extrapolation in Image Diffusion Transformers

TL;DR

UltraImage tackles resolution extrapolation in image diffusion transformers by (1) identifying RoPE-based repetition as a mid-band frequency effect and applying recursive dominant frequency correction to keep the dominant frequency within a single period during extrapolation, and (2) diagnosing quality degradation as attention dilution and solving it with entropy-guided adaptive attention concentration, implemented efficiently via a Triton-based kernel. The method yields consistent improvements over prior work on Flux and Qwen-Image across direct, guided, and guided-view extrapolation, and demonstrates extreme extrapolation up to 6K×6K from 1328p without low-resolution guidance. These contributions enable high-fidelity, coherent ultra-resolution image generation, with practical impact for large-format media and high-detail simulations. The work advances understanding of frequency-domain effects in RoPE and introduces a per-head, entropy-aware attention sharpening mechanism that preserves global structure while enhancing local details.

Abstract

Recent image diffusion transformers achieve high-fidelity generation, but struggle to generate images beyond these scales, suffering from content repetition and quality degradation. In this work, we present UltraImage, a principled framework that addresses both issues. Through frequency-wise analysis of positional embeddings, we identify that repetition arises from the periodicity of the dominant frequency, whose period aligns with the training resolution. We introduce a recursive dominant frequency correction to constrain it within a single period after extrapolation. Furthermore, we find that quality degradation stems from diluted attention and thus propose entropy-guided adaptive attention concentration, which assigns higher focus factors to sharpen local attention for fine detail and lower ones to global attention patterns to preserve structural consistency. Experiments show that UltraImage consistently outperforms prior methods on Qwen-Image and Flux (around 4K) across three generation scenarios, reducing repetition and improving visual fidelity. Moreover, UltraImage can generate images up to 6K*6K without low-resolution guidance from a training resolution of 1328p, demonstrating its extreme extrapolation capability. Project page is available at \href{https://thu-ml.github.io/ultraimage.github.io/}{https://thu-ml.github.io/ultraimage.github.io/}.

Figures (18)

• Figure 1: Generated results of UltraImage. Starting from the base Qwen-Image model trained at $1328\text{p}$ resolution, UltraImage can generate high-quality images up to $6\text{K} \times 6\text{K}$ without any low-resolution guidance, demonstrating its extreme extrapolation capability. All prompts used in this paper are provided in the Appendix.
• Figure 2: Challenges of resolution extrapolation in image diffusion transformers. Top row: Flux (training resolution with $2048^2$). Bottom row: Qwen-Image (training resolution $1328\text{p}$). Both models exhibit typical failure modes: content repetition and quality degradation at higher resolutions.
• Figure 3: Cause of content repetition.Left: (a) Height extrapolation baselines. (b) High-frequency interpolation blurs local textures. (c) The mid-band dominant frequency, whose period aligns with the training height $h$, governs global structure and introduces repetition. (d) Low-frequency components have minimal effect. Right: Validation: repetition appears when the extrapolated height $H$ (e) exceeds the dominant period $T_k^h$, and disappears when $H \le T_k^h$ (f,g).
• Figure 4: Failure modes of existing positional extrapolation. NTK suffers from repetition during extrapolation, while PI reduces it but causes over-smoothed textures. In contrast, our method mitigates repetition without sacrificing image fidelity by correctly identifying and adjusting the relevant frequency components.
• Figure 5: Cause of quality degradation. Previous interpolation of a single high frequency (HF) in Fig. \ref{['fig:each frequency']}b leads to similar quality loss during extrapolation. (a) Comparing attention maps (b) pre- and (c) post-interpolation, the distribution becomes significantly flattened. Sharpening the attention in (d) restores the lost details, confirming that the degradation originates from reduced attention concentration.