Breaking the Static: Dynamic Text Conditioning for Diverse Image Generation
Dynamic Semantic Interpolation aligns text conditioning with the denoising process to improve generation diversity.
VISUAL GENERATION · MULTIMODAL UNDERSTANDING
Visual generation.
Multimodal understanding.
I am a Ph.D. student (2023–present) in the School of Electrical and Information Engineering at Tianjin University, supervised by Prof. An-An Liu and Prof. Lanjun Wang. Alongside my doctoral research, I interned with the DreamX team at Amap, Alibaba Group, and am now interning with the Bailing Multimodal Team at Ant Group through the PlanA program.
My research focuses on visual generation and multimodal understanding, with interests in image and video synthesis, controllable generation, and generative model safety.
Interested in my research? chenruidong@tju.edu.cn
News
Past 12 monthsOur paper “Breaking the Static: Dynamic Text Conditioning for Diverse Image Generation” is accepted to NeurIPS 2026.
Our paper “What Concepts Lie Within? Detecting and Suppressing Risky Content in Diffusion Transformers” is accepted to ACM CCS 2026.
Our paper “Cognitive Distillation for Information Forensics: Towards Improved Hateful Meme Detection” is accepted to KDD 2026.
Show 4 more updates
Our paper “Layer-wise Instance Binding for Regional and Occlusion Control in Text-to-Image Diffusion Transformers” is accepted to CVPR 2026.
Our paper “Group Relative Attention Guidance for Image Editing” is accepted to CVPR 2026 Findings.
Our paper “Eevee: Towards Close-up High-resolution Video-based Virtual Try-on” is accepted to CVPR 2026 Findings.
Our paper “T2I-RiskyPrompt: A Benchmark for Safety Evaluation, Attack, and Defense on Text-to-Image Model” is accepted to AAAI 2026.
Selected Publications 6
View allFirst-author, co-first-author, and first-student-author work.
Layer-wise Instance Binding for Regional and Occlusion Control in Text-to-Image Diffusion Transformers
Training-free regional and occlusion control for text-to-image diffusion transformers.
TRCE: Towards Reliable Malicious Concept Erasure in Text-to-Image Diffusion Models
Reliable malicious concept erasure for safer text-to-image diffusion models.
T2TD: Text-3D generation model based on prior knowledge guidance
Guiding text-to-3D generation with prior knowledge.
CompCraft: Foreground-driven Image Synthesis with Customized Layouts
Composing foreground subjects into images with customized spatial layouts.
AnyScene: Customized Image Synthesis with Composited Foreground
Customized image synthesis that brings composited foreground subjects into a coherent scene.
All Publications 14
Google ScholarA complete list, including the selected work above.
14 papers
2026
Breaking the Static: Dynamic Text Conditioning for Diverse Image Generation
What Concepts Lie Within? Detecting and Suppressing Risky Content in Diffusion Transformers
2025
No matching publications. Try another title, author, venue, or year.