We are delighted to announce that 1 paper from IRIS Lab (Prof. Hak Gu Kim) has been accepted to NeurIPS 2026.

Title: 

Seeing Speech: Learning Visible Articulatory Dynamics for Speech-Driven 3D Facial Animation

Authors: 

Hyung Kyu Kim, Byungchan Hwang, Hak Gu Kim

Abstract: 

Speech-driven facial animation can reproduce 3D vertex positions without fully capturing visible speech articulation, because acoustic signals can correspond to multiple facial motions. This work models lip spreading, opening, and protrusion as complementary directional movements. Its Speech-Articulatory Memory retrieves phonetic-context-aware motion patterns, and Topology-aware Articulatory Composition combines them into coherent 3D mesh animations. Evaluations on VOCASET and TFHP report improved reconstruction and visible lip dynamics, while a user study favors the resulting lip synchronization and realism.