Mosaic-SDF for 3D Generative Models

Explore the innovative Mosaic-SDF representation for 3D generative models in this comprehensive talk by Lior Yariv from Valence Labs. Dive into the design principles of effective shape representations and discover how Mosaic-SDF addresses these principles by approximating the Signed Distance Function using local grids near shape boundaries. Learn about the advantages of this representation, including fast computation, parameter efficiency, and compatibility with Transformer-based architectures. Follow the implementation process from preprocessing to generation using flow matching techniques. Examine the results of class-conditioned generation with the 3D Warehouse dataset and text-to-3D generation using a large caption-shape pair dataset. Gain insights into the future of 3D shape generation through the Q&A session at the end of this informative presentation.

- Intro + Background
- Representation Design Principles
- Mosaic-SDF Representation
- Mosaic-SDF Preprocess
- Mosaic-SDF Generation with Flow Matching
- Results
- Q+A