3D Dance Generation from Music: Model Fine-Tuning

In this research, I fine-tuned and debugged a generative AI model capable of transforming music into dynamic 3D dance motions. To create a robust dataset, I meticulously selected, trimmed, and adjusted the tempo of various music tracks to align seamlessly with existing video content. This project not only showcases my technical skills but also earned me the 2023 College Student Research Scholarship, highlighting my commitment to advancing creative applications of AI in computer vision.
Github Repo Link: github.com/CaslowChien/EDGE

Research Results

Fine-tuned a 3D dance video generation model with PyTorch to specialize in breakdancing, expanding the dataset from 1,269 to 1,435 motion-music pairs with self-edited clips, resulting in an 18% increase in votes in a performance blind test.
Awarded the 2023 College Student Research Scholarship
Accepted to present an oral paper at The International Workshop on Advanced Image Technology 2025.

Final Presentation

Generative AI

Implemented a diffusion model that combines audio and video, significantly expanding my understanding of computer vision AI models in real-world scenarios.

Information Research

The complexity of this model and its dataset has made me more familiar with understanding the essence of intricate model structures.

Model & Validation

I applied model validation techniques, including methods like cross-validation.