# First-Author Papers

ControlScene: Controllable Text-to-3D Scene Generation via Structured Layout Priors
Ruyi Zhang, et al.
CAAI Trans. · Accepted 3D Generation LLM

• The LayoutVerse-20K dataset with 20,000 manually annotated scenes (prompts, layouts, scene graphs, 3D scenes).
• An LLM-driven structured layout generation framework and two new metrics: category plausibility and layout plausibility.

Fine-Grained Cross-Modal Alignment for Unlinked Visual References in Scientific Papers
Ruyi Zhang, et al.
Ready for Submission MLLM Document Understanding

• Count-Curriculum, a three-stage curriculum fine-tuning strategy that trains MLLMs from easy to hard figure-matching samples; reaches 75.70% accuracy on the number-free figure-matching task.
• A benchmark of 21,985 citation-figure pairs from 1,165 arXiv papers across 30+ disciplines.

# Co-Authored Papers

Generalize LMMs to Versatile Visual Modalities via Fabricated Modality Synthesis
S. Yuan, Y. Li, Ruyi Zhang, et al.
ECCV 2026 · Accepted LMM Generalization

• My role: design of the modality synthesis pipeline and zero-shot generalization experiments.

ScrollScape: Unlocking 32K Image Generation With Video Diffusion Priors
H. Yu, Y. Zhang, D. Di, Ruyi Zhang, et al.
ECCV 2026 · Accepted Diffusion Image Generation

• My role: ultra-high-resolution image generation and validation of the video diffusion prior module.

Task-Uncertainty-Aware Video Restoration for Time-varying Unknown Degradations
Wenrui Li, Hongtao Chen, Ruyi Zhang, Zhe Yang, Wangmeng Zuo
IEEE TMM · Accepted Video Restoration

• My role: validation of the spatio-temporal uncertainty regularization and quantitative comparison on multiple benchmarks.

Hierarchical Semantic-Constrained Heterogeneous Graph for Audio-Visual Event Localization
Zhe Yang, Ruyi Zhang, Hongtao Chen, Wenrui Li, Hengyu Man
IEEE TIP · Under Review Audio-Visual

• My role: model validation, hyperbolic-space visualization and curation of the OV-AVEBench dataset.