Tuning-Free Latent Diffusion Models for Ultrahigh-Resolution Image Editing
Published in IEEE Transactions on Neural Networks and Learning Systems, 1-15, 2026
Wanglong Lu, Lingming Su, Kaijie Shi, Minglun Gong, Xiaogang Jin, Hanli Zhao, Xianta Jiang
Brief description:
UltraDiffEdit is a tuning-free framework for ultra-high-resolution real-image editing. It extends pre-trained latent diffusion models to high-resolution inputs without additional training or per-image fine-tuning, using a coarse-to-fine progressive editing process to preserve unedited regions while refining generated content.
Highlights:
- Multi-scale progressive editing with an encode-diffuse-denoise-decode-blend loop.
- Multi-patch encoding for preserving high-resolution edited and unedited details in latent space.
- Global-local consistency denoising for reducing boundary artifacts and preserving background content.
- Patch-based hybrid sampling to combine local, intermediate, and global context; demonstrated up to 8K resolution on a single NVIDIA RTX 3090 GPU.
[paper] [pdf] [html] [doi] [github]
Recommended citation:
@article{lu2026tuningfree,
title={Tuning-Free Latent Diffusion Models for Ultrahigh-Resolution Image Editing},
author={Lu, Wanglong and Su, Lingming and Shi, Kaijie and Gong, Minglun and Jin, Xiaogang and Zhao, Hanli and Jiang, Xianta},
journal={IEEE Transactions on Neural Networks and Learning Systems},
pages={1--15},
doi={10.1109/TNNLS.2026.3707463},
year={2026}
}