Tianyi Wei
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Tianyi Wei has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
- Nenghai Yu
University of Science and Technology of China, Hefei, China
9 shared publications - Weiming Zhang
University of Science and Technology of China, Hefei, China
8 shared publications - Wenbo Zhou
School of Cyber Science and Technology, University of Science and Technology of China, China
6 shared publications - Dongdong Chen
Microsoft GenAI, Redmond, WA, USA
5 shared publications - Gang Hua
Dolby Laboratories, Bellevue, WA, USA
4 shared publications - Jing Liao
City University of Hong Kong
4 shared publications
Latest Publications
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 3235-3249
Token Buncher : Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 3235-3249
Open AccessToken Buncher : Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning
- Open Access
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 6643-6654
HybridDet: Redistribution-Robust Visual Content Provenance for Information Credibility
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 6643-6654
Open AccessHybridDet: Redistribution-Robust Visual Content Provenance for Information Credibility
- 10 September 2026
- Published by Springer Nature
- p. 586-605
Towards Geometry-Grounded Dense Semantic Matching with VGGT Priors
- 10 September 2026
- Published by Springer Nature
- p. 586-605
Towards Geometry-Grounded Dense Semantic Matching with VGGT Priors
- Top Cited Papers
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (7) , 7692-7706
Enhancing MMDiT-Based Text-to-Image Models for Similar Subject Generation
Representing the cutting-edge technique of text-to-image models, the latest Multimodal Diffusion Transformer (MMDiT) largely mitigates many generation issues existing in previous models. However, we discover that it still suffers from subject neglect or mixing when the input text prompt contains multiple subjects of similar semantics or appearance. We identify three possible ambiguities within the MMDiT architecture that cause this problem: Inter-block Ambiguity, Text Encoder Ambiguity, and Semantic Ambiguity. To address these issues, we propose to repair the ambiguous latent on-the-fly by test-time optimization at early denoising steps. In detail, we design three loss functions: Block Alignment Loss, Text Encoder Alignment Loss, and Overlap Loss, each tailored to mitigate these ambiguities. Despite significant improvements, we observe that semantic ambiguity persists when generating multiple similar subjects, as the guidance provided by overlap loss is not explicit enough. Therefore, we further propose Overlap Online Detection and Back-to-Start Sampling Strategy to alleviate the problem. Experimental results on a newly constructed challenging dataset of similar subjects validate the effectiveness of our approach, showing superior generation quality and much higher success rates over existing methods. The consistent and substantial improvements observed across multiple MMDiT based text-to-image models such as SD3, SD3.5 and FLUX provide strong evidence of the general applicability of our method.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (7) , 7692-7706
Top Cited PapersEnhancing MMDiT-Based Text-to-Image Models for Similar Subject Generation
Representing the cutting-edge technique of text-to-image models, the latest Multimodal Diffusion Transformer (MMDiT) largely mitigates many generation issues existing in previous models. However, we discover that it still suffers from subject neglect or mixing when the input text prompt contains multiple subjects of similar semantics or appearance. We identify three possible ambiguities within the MMDiT architecture that cause this problem: Inter-block Ambiguity, Text Encoder Ambiguity, and Semantic Ambiguity. To address these issues, we propose to repair the ambiguous latent on-the-fly by test-time optimization at early denoising steps. In detail, we design three loss functions: Block Alignment Loss, Text Encoder Alignment Loss, and Overlap Loss, each tailored to mitigate these ambiguities. Despite significant improvements, we observe that semantic ambiguity persists when generating multiple similar subjects, as the guidance provided by overlap loss is not explicit enough. Therefore, we further propose Overlap Online Detection and Back-to-Start Sampling Strategy to alleviate the problem. Experimental results on a newly constructed challenging dataset of similar subjects validate the effectiveness of our approach, showing superior generation quality and much higher success rates over existing methods. The consistent and substantial improvements observed across multiple MMDiT based text-to-image models such as SD3, SD3.5 and FLUX provide strong evidence of the general applicability of our method.
- 5 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
Rank-based No-reference Quality Assessment for Face Swapping
Face swapping has emerged as a significant research area in generative models due to rapid technological advancements. Current image quality assessment metrics for face swapping images depend on comparing facial attributes between fake and real images (e.g., target faces and source faces). However, obtaining such corresponding real images is often impractical. To address this challenge, we propose a no-reference image quality assessment method for face swapping. We observed that consistency among errors in various facial attributes can indicate overall image quality. Based on this insight, we created a large-scale face swapping dataset with multi-attribute ranking labels to produce quality ranking labels for objective attribute errors. We then trained a Siamese network to learn consistency patterns among these errors, demonstrating through experiments that these patterns align with human subjective judgments. Our method achieved state-of-the-art results in both coarse- and fine-grained tests. Finally, we applied this as a loss function to enhance existing face swapping models, leading to improvements in expression and pose attributes. Extensive experiments confirm the superiority of our method over existing general no-reference image quality assessment metrics and the latest metric for facial image quality assessment, making it well-suited for evaluating face swapping images in real-world scenarios.
- 5 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
Rank-based No-reference Quality Assessment for Face Swapping
Face swapping has emerged as a significant research area in generative models due to rapid technological advancements. Current image quality assessment metrics for face swapping images depend on comparing facial attributes between fake and real images (e.g., target faces and source faces). However, obtaining such corresponding real images is often impractical. To address this challenge, we propose a no-reference image quality assessment method for face swapping. We observed that consistency among errors in various facial attributes can indicate overall image quality. Based on this insight, we created a large-scale face swapping dataset with multi-attribute ranking labels to produce quality ranking labels for objective attribute errors. We then trained a Siamese network to learn consistency patterns among these errors, demonstrating through experiments that these patterns align with human subjective judgments. Our method achieved state-of-the-art results in both coarse- and fine-grained tests. Finally, we applied this as a loss function to enhance existing face swapping models, leading to improvements in expression and pose attributes. Extensive experiments confirm the superiority of our method over existing general no-reference image quality assessment metrics and the latest metric for facial image quality assessment, making it well-suited for evaluating face swapping images in real-world scenarios.
- 21 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (6) , 6407-6424
Unifying Multi-Modal Hair Editing via Proxy Feature Blending
Hair editing is a long-standing problem in computer vision that demands both fine-grained local control and intuitive user interactions across diverse modalities. Despite the remarkable progress of GANs and diffusion models, existing methods still lack a unified framework that simultaneously supports arbitrary interaction modes (e.g., text, sketch, mask, and reference image) while ensuring precise editing and faithful preservation of irrelevant attributes. In this work, we introduce a novel paradigm that reformulates hair editing as proxy-based hair transfer. Specifically, we leverage the dense and semantically disentangled latent space of StyleGAN for precise manipulation and exploit its feature space for disentangled attribute preservation, thereby decoupling the objectives of editing and preservation. Our framework unifies different modalities by converting editing conditions into distinct transfer proxies, whose features are seamlessly blended to achieve global or local edits. Beyond 2D, we extend our paradigm to 3D-aware settings by incorporating EG3D and PanoHead, where we propose a multi-view boosted hair feature localization strategy together with 3D-tailored proxy generation methods that exploit the inherent properties of 3D-aware generative models. Extensive experiments demonstrate that our method consistently outperforms prior approaches in editing effects, attribute preservation, visual naturalness, and multi-view consistency, while offering unprecedented support for multimodal and mixed-modal interactions.
- 21 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (6) , 6407-6424
Unifying Multi-Modal Hair Editing via Proxy Feature Blending
Hair editing is a long-standing problem in computer vision that demands both fine-grained local control and intuitive user interactions across diverse modalities. Despite the remarkable progress of GANs and diffusion models, existing methods still lack a unified framework that simultaneously supports arbitrary interaction modes (e.g., text, sketch, mask, and reference image) while ensuring precise editing and faithful preservation of irrelevant attributes. In this work, we introduce a novel paradigm that reformulates hair editing as proxy-based hair transfer. Specifically, we leverage the dense and semantically disentangled latent space of StyleGAN for precise manipulation and exploit its feature space for disentangled attribute preservation, thereby decoupling the objectives of editing and preservation. Our framework unifies different modalities by converting editing conditions into distinct transfer proxies, whose features are seamlessly blended to achieve global or local edits. Beyond 2D, we extend our paradigm to 3D-aware settings by incorporating EG3D and PanoHead, where we propose a multi-view boosted hair feature localization strategy together with 3D-tailored proxy generation methods that exploit the inherent properties of 3D-aware generative models. Extensive experiments demonstrate that our method consistently outperforms prior approaches in editing effects, attribute preservation, visual naturalness, and multi-view consistency, while offering unprecedented support for multimodal and mixed-modal interactions.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28, 12304-12315
StructDiff: A Structure-Preserving and Spatially Controllable Diffusion Model for Single-Image Generation
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28, 12304-12315
StructDiff: A Structure-Preserving and Spatially Controllable Diffusion Model for Single-Image Generation
- Open Access
- 14 December 2025
- Published by Association for Computing Machinery (ACM)
- p. 1-11
Bokeh Diffusion: Defocus Blur Control in Text-to-Image Diffusion Models
- 14 December 2025
- Published by Association for Computing Machinery (ACM)
- p. 1-11
Open AccessBokeh Diffusion: Defocus Blur Control in Text-to-Image Diffusion Models
- 26 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 13055-13070
Antiano: A Series of Attacks Exploiting Vulnerabilities in Deep Face Anonymization Algorithms
Deep learning-based techniques for facial anonymization strive to protect identity details while preserving image usability. We present Antiano, the first work to reveal possible security risks associated with deep learning-based facial anonymization algorithms. By evaluating the entire anonymization process, we introduce a framework comprising four attack types: Anonymization Deactivation Attack, ID Destruction Attack, Downstream Task Attack, and Attribute Attack. These attacks aim to undermine anonymization effectiveness, mislead anonymized outputs and downstream tasks, or modify facial attributes in subtle ways. For each scenario, we develop novel strategies that craft perturbations tailored to a user’s face and transferable across different keys and models. Experimental findings reveal that Antiano achieves significant attack effectiveness against various reversible anonymization models, highlighting security vulnerabilities in current deep learning anonymization algorithms. This paper also distinguishes Antiano from traditional adversarial attacks and provides security recommendations for the future development of facial anonymization technologies.
- 26 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 13055-13070
Antiano: A Series of Attacks Exploiting Vulnerabilities in Deep Face Anonymization Algorithms
Deep learning-based techniques for facial anonymization strive to protect identity details while preserving image usability. We present Antiano, the first work to reveal possible security risks associated with deep learning-based facial anonymization algorithms. By evaluating the entire anonymization process, we introduce a framework comprising four attack types: Anonymization Deactivation Attack, ID Destruction Attack, Downstream Task Attack, and Attribute Attack. These attacks aim to undermine anonymization effectiveness, mislead anonymized outputs and downstream tasks, or modify facial attributes in subtle ways. For each scenario, we develop novel strategies that craft perturbations tailored to a user’s face and transferable across different keys and models. Experimental findings reveal that Antiano achieves significant attack effectiveness against various reversible anonymization models, highlighting security vulnerabilities in current deep learning anonymization algorithms. This paper also distinguishes Antiano from traditional adversarial attacks and provides security recommendations for the future development of facial anonymization technologies.