Lee, MyeongjinLibao, Emmanuel IanLee, Sung-HeeChristie, MarcHan, Ping-HsuanLin, Shih-SyunPietroni, NicoSchneider, TeseoTsai, Hsin-RueyWang, Yu-ShuenZhang, Eugene2025-10-072025-10-072025978-3-03868-295-0https://doi.org/10.2312/pg.20251259https://diglib.eg.org/handle/10.2312/pg20251259We present a hybrid retrieval-regression framework for motion-driven garment animation leveraging a shared discrete codebook. Our method targets the challenge of animating loose-fitting garments, whose dynamic behaviors exhibit high variability and less direct correlation with body motion-making them difficult to handle with conventional example-based approaches that assume tightly coupled motion-garment relationships. To address this, we project both motion and garment animation clips into a shared discrete codebook via Gumbel-Softmax-based quantization, allowing them to be aligned in a semantically consistent space where cross-retrieval can be performed using simple distance metrics. During inference, we adaptively switch between retrieval and regression based on the confidence derived from the codebook probability distribution, allowing the system to remain robust in the presence of ambiguous or unseen motions. We leverage a pre-trained mesh autoencoder to obtain garment latents that preserve local geometric structure, enabling smoother transitions and more geometrically consistent interpolation between retrieved and regressed animation segments efficiently. Experimental results demonstrate that our approach improves the accuracy and plausibility of garment animation for complex garments under diverse motion inputs, while maintaining robustness to unseen scenarios and achieving low simulation error for high-quality garment animation.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Animation; Learning latent representations; Discrete space searchComputing methodologies → AnimationLearning latent representationsDiscrete space search10.2312/pg.2025125910 pages