| --- |
| license: mit |
| language: |
| - en |
| --- |
| |
| <h1 align='center'>ELBO-T2IAlign: A Generic ELBO-Based Method for Calibrating Pixel-level Text-Image Alignment in Diffusion Models</h1> |
| <p align="center"> <span style="color:#137cf3; font-family: Gill Sans">Qin Zhou</span>, <span style="color:#137cf3; font-family: Gill Sans">Zhiyang Zhang</span>, <span style="color:#137cf3; font-family: Gill Sans">Jinglong Wang</span>, <span style="color:#137cf3; font-family: Gill Sans">Xiaobin Li</span>, <span style="color:#137cf3; font-family: Gill Sans">Jing Zhang</span><sup>*</sup>, <span style="color:#137cf3; font-family: Gill Sans">Qian Yu</span>, <span style="color:#137cf3; font-family: Gill Sans">Lu Sheng</span>, <span style="color:#137cf3; font-family: Gill Sans">Dong Xu</span> <br> |
| <span style="font-size: 16px">Beihang University</span>, <span style="font-size: 16px">University of Hong Kong</span></p> |
| |
| <div align="center"> |
| <a href="https://vcg-team.github.io/elbo-t2ialign-webpage/"><img src="https://img.shields.io/static/v1?label=elbo-t2ialign&message=Project&color=purple"></a> |
| <a href="https://arxiv.org/abs/2506.09740"><img src="https://img.shields.io/static/v1?label=Paper&message=Arxiv&color=red&logo=arxiv"></a> |
| <a href="https://github.com/VCG-team/elbo-t2ialign"><img src="https://img.shields.io/static/v1?label=Code&message=Github&color=blue&logo=github"></a> |
| <a href="https://huggingface.co/datasets/Matrix53/elbo-t2ialign"><img src="https://img.shields.io/static/v1?label=Dataset&message=HuggingFace&color=yellow&logo=huggingface"></a> |
| </div> |
| |
| ## Abstract |
| |
| Diffusion models excel at image generation. Recent studies have shown that these models not only generate high-quality images but also encode text-image alignment information through attention maps or loss functions. This information is valuable for various downstream tasks, including segmentation, text-guided image editing, and compositional image generation. However, current methods heavily rely on the assumption of perfect text-image alignment in diffusion models, which is not the case. In this paper, we propose using zero-shot referring image segmentation as a proxy task to evaluate the pixel-level image and class-level text alignment of popular diffusion models. We conduct an in-depth analysis of pixel-text misalignment in diffusion models from the perspective of training data bias. We find that misalignment occurs in images with small-sized, occluded, or rare object classes. Therefore, we propose ELBO-T2IAlign—a simple yet effective method to calibrate pixel-text alignment in diffusion models based on the evidence lower bound (ELBO) of likelihood. ELBO-T2IAlign is training-free and generic: it requires no additional annotations, model retraining, or architectural modifications, and it can be directly applied to different diffusion backbones. Extensive experiments on zero-shot referring image segmentation, text-guided image editing, and compositional image generation verify that the proposed calibration improves pixel-text alignment across complementary downstream tasks. |
| |
| ## Details |
| |
| This repository contains all datasets used in our paper, including COCO, VOC, Context... |
| |
| Only validation sets are included, which cost about 7G memory. |
| ```bash |
| # unzip command |
| cat dataset.tar.gz.* > dataset.tar.gz |
| tar -xzf dataset.tar.gz |
| ``` |
| |
| ## Citation |
| ```bibtex |
| @article{zhou2025elbo, |
| title={ELBO-T2IAlign: A Generic ELBO-Based Method for Calibrating Pixel-level Text-Image Alignment in Diffusion Models}, |
| author={Zhou, Qin and Zhang, Zhiyang and Wang, Jinglong and Li, Xiaobin and Zhang, Jing and Yu, Qian and Sheng, Lu and Xu, Dong}, |
| journal={arXiv preprint arXiv:2506.09740}, |
| year={2025} |
| } |
| ``` |
