HRDiT: Training-Free High-Resolution Image Generation with Off-the-Shelf Diffusion Transformer Models

2026-08-07

AI Summary

This work addresses two key challenges in high-resolution image generation with off-the-shelf Diffusion Transformer (DiT) models: spatial disorganization and low generation efficiency. The authors propose a training-free adaptation method that, for the first time, enables pre-trained DiT models to generate high-resolution images without any additional training or parameter updates. By keeping the model parameters frozen, the approach significantly enhances spatial coherence and structural integrity of generated images while substantially reducing inference time. Experimental results demonstrate that this strategy effectively overcomes the longstanding reliance on U-Net architectures for high-resolution synthesis, thereby opening a new pathway for leveraging DiTs in efficient, high-quality image generation.

Abstract

Training-free text-to-high-resolution image generation has recently attracted growing research attention. However, existing studies on this task primarily focus on adapting off-the-shelf U-Net-based diffusion models to high resolutions, with limited progress on adapting off-the-shelf Diffusion Transformer (DiT) models despite their strong text-to-image generation capabilities at limited resolutions. In this work, we find two key challenges particularly hindering the application of off-the-shelf DiT models for high-resolution image synthesis in a training-free manner, namely, spatial disorder and long generation time. To address these challenges, we propose a novel method tailored to adapt off-the-shelf DiT models for high-resolution image synthesis. Extensive experiments show the efficacy of our method. Our code is available at: https://github.com/zylwithxy/HRDiT.

Problem

• high-resolution image generation
• training-free
• Diffusion Transformer
• text-to-image synthesis
• off-the-shelf models

Innovation

• training-free
• high-resolution image generation
• Diffusion Transformer
• spatial disorder
• efficient generation