Shu-Yuan Tu

A novel event-centric evaluation framework, AVC-Reward, finally provides a nuanced assessment of audio-video generation beyond holistic metrics, revealing fine-grained cross-modal failures.

Tool-calling and multi-role prompting don't make a visual generation pipeline truly agentic—autonomy fundamentally requires shifting controller authority from prompt formatting to closed-loop visual self-correction and cross-task memory.

VA-Judger transforms reward modeling by integrating human preference feedback, leading to coherent and high-quality video-audio generation that outperforms traditional metrics.

Forget task-specific fine-tuning: PSD aligns text-to-3D generation with human preferences using only pre-trained 2D reward models.

Emotion-LLaMAv2 leaps beyond its predecessor by ditching face detectors and embracing a unified architecture that actually understands emotion, thanks to a novel pre-fusion module and curriculum tuning.