Inf-DiT: Upsampling Any-Resolution Image with Memory-Efficient Diffusion Transformer.
Abstract
Diffusion models have shown remarkable performance in image generation in recent years. However, due to a quadratic increase in memory during generating ultra-high-resolution images (e.g. 4096 × 4096), the resolution of generated images is often limited to 1024×1024. In this work. we propose a unidirectional block attention mechanism that can adaptively adjust the memory overhead during the inference process and handle global dependencies. Building on this module, we adopt the DiT structure for upsampling and develop an infinite super-resolution model capable of upsampling images of various shapes and resolutions. Comprehensive experiments show that our model achieves SOTA performance in generating ultra-high-resolution images in both machine and human evaluation. Compared to commonly used UNet structures, our model can save more than 5× memory when generating 4096 × 4096 images. The project URL is https://github.com/THUDM/Inf-DiT.
Cite
Text
Yang et al. "Inf-DiT: Upsampling Any-Resolution Image with Memory-Efficient Diffusion Transformer.." Proceedings of the European Conference on Computer Vision (ECCV), 2024. doi:10.1007/978-3-031-73010-8_9Markdown
[Yang et al. "Inf-DiT: Upsampling Any-Resolution Image with Memory-Efficient Diffusion Transformer.." Proceedings of the European Conference on Computer Vision (ECCV), 2024.](https://mlanthology.org/eccv/2024/yang2024eccv-infdit/) doi:10.1007/978-3-031-73010-8_9BibTeX
@inproceedings{yang2024eccv-infdit,
title = {{Inf-DiT: Upsampling Any-Resolution Image with Memory-Efficient Diffusion Transformer.}},
author = {Yang, Zhuoyi and Jiang, Heyang and Hong, Wenyi and Teng, Jiayan and Zheng, Wendi and Dong, Yuxiao and Ding, Ming and Tang, Jie},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
year = {2024},
doi = {10.1007/978-3-031-73010-8_9},
url = {https://mlanthology.org/eccv/2024/yang2024eccv-infdit/}
}