UltraImage: Rethinking Resolution Extrapolation in Image Diffusion Transformers
This addresses a key limitation in high-resolution image generation for AI and creative applications, representing a strong specific gain rather than a foundational breakthrough.
The paper tackled the problem of image diffusion transformers struggling to generate images beyond training scales, which causes content repetition and quality degradation, by introducing UltraImage, a framework that reduces repetition and improves visual fidelity, enabling generation up to 6K*6K from a 1328p training resolution.
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/}.