Shape Generation via Weight Space Learning

Maximilian Plattner, Arturs Berzins, Johannes Brandstetter

Published: 05 Mar 2025, Last Modified: 05 Mar 2025ICLR 2025 Workshop Weight Space Learning PosterEveryoneRevisionsBibTeXCC BY 4.0
Track: tiny / short paper (up to 4 pages)
Keywords: weight space learning, shape generation, neural field, diffusion, foundation model, interpretability, fine-tuning, downstream task, foundation model, 3D shape-generative foundation model
TL;DR: We show how treating a 3D shape-generative foundation model’s weight space as a data modality uncovers low-dimensional submanifolds that enable selective control of global topology or local geometry, even under scarce data.
Submission Number: 22