Research outputs

Selected Publications

Representative work in visual quality assessment, visual attention, and reliable vision-language models.

Diagram of Q-Debias, which exposes perception bias with a distorted conditional image and mitigates it through quality-aware prompting.
Q-Debias framework

AAAI 2026 Best Poster Paper

Mitigating Perception Bias: A Training-Free Approach to Enhance LMM for Image Quality Assessment

Baoliang Chen, Siyi Pan, Dongxu Wu, Liang Xie, Xiangjie Sui, Lingyu Zhu, Hanwei Zhu

We expose semantic bias in large multimodal models with controlled degradations, then use a training-free prompting scheme to make image-quality judgments more perceptually reliable.

PaperProject & code

The same vision-language model answering a visual question for a clean image and a defocus-corrupted image.
Corruption-induced reliability failure

arXiv · 2025

Diagnosing Corruption-Induced Reliability Failures in Vision-Language Models

Xiangjie Sui, Songyang Li, Hanwei Zhu, Baoliang Chen, Yuming Fang, Xin Sun

We benchmark how visual corruptions alter vision-language model responses and introduce a discriminative robustness alignment metric to expose reliability failures in perceptual tasks.

PaperCode

Diagram of the generative scanpath representation converter and quality evaluator for omnidirectional images.
Generative scanpath representation

IEEE Transactions on Image Processing · 2025

Perceptual Quality Assessment of 360° Images Based on Generative Scanpath Representation

Xiangjie Sui, Hanwei Zhu, Xuelin Liu, Yuming Fang, Shiqi Wang, Zhou Wang

We model the viewing experience of 360° images as generative scanpath sequences, focusing quality assessment on the locally attended regions that drive perceptual judgments.

PaperCode

Two-alternative forced choice examples used to evaluate image quality assessment by large multimodal models.
2AFC probing for LMMs

IEEE Transactions on Circuits and Systems for Video Technology · 2024

2AFC Prompting of Large Multimodal Models for Image Quality Assessment

Hanwei Zhu, Xiangjie Sui, Baoliang Chen, Xuelin Liu, Peilin Chen, Yuming Fang, Shiqi Wang

We adapt the two-alternative forced-choice protocol from vision science to probe and quantify the image-quality assessment capability of large multimodal models.

PaperCode

The ScanDMM framework with state initialization, transition, and emission functions for 360-degree scanpath prediction.
Proposed ScanDMM framework

CVPR 2023

ScanDMM: A Deep Markov Model of Scanpath Prediction for 360° Images

Xiangjie Sui, Yuming Fang, Hanwei Zhu, Shiqi Wang, Zhou Wang

ScanDMM learns the temporal dynamics of visual attention to predict realistic scanpaths in 360° scenes, beyond static saliency maps and unconstrained generative trajectories.

PaperCode

Framework that samples a scanpath-conditioned video sequence and pools frame-level quality scores for omnidirectional image assessment.
Proposed omnidirectional IQA framework

IEEE Transactions on Visualization and Computer Graphics · 2022

Perceptual Quality Assessment of Omnidirectional Images as Moving Camera Videos

Xiangjie Sui, Kede Ma, Yiru Yao, Yuming Fang

We formulate the perceived quality of omnidirectional images through moving-camera videos that reflect how viewers explore an immersive scene.

PaperCode

For the complete and most current publication record, see Google Scholar.