Ind3D: Enforcing Inductive Bias in 3D Generation

from Geometric, Physical, Topological, and Functional Perspectives

CVPR 2025

📍 Location: Davidson C3

đź“… Date: June 12, 1:00 PM - 5:30 PM

Abstract

The current era of generative AI thrives on scaling data and models, yet fundamental challenges remain, particularly in preserving geometric, physical, topological, and functional priors in 3D generation. Unlike video generation models, such as OpenAI SORA, where shortcomings in physics or geometry do not critically undermine the method, 3D generation requires strict compliance with rich priors. Man-made objects, for instance, should exhibit strong primitive shapes, physical stability, and usability through human interaction. Existing approaches to address these issues focus on either network designs, such as equivariant networks, TutteNet, and CoFie, or regularization losses to enforce geometric structures, physical stability, and topological generalization. However, these efforts are often fragmented, focusing on specific priors, and are overshadowed by the emphasis on scaling datasets and model architectures.
This workshop aims to unite researchers to advance the integration of inductive biases in 3D generative models, providing tools that are simple, efficient, and impactful. By addressing the challenges of modeling inductive biases—which require interdisciplinary knowledge and often incur training costs—this workshop seeks to foster a community dedicated to these methods. As generative AI faces limits in training data and model scalability, incorporating inductive biases offers a path forward, enabling smaller models with reduced data requirements. While the focus is on 3D inductive biases due to their abundance of priors, the workshop's relevance extends to the broader computer vision community. Key topics include embedding geometrical, physical, topological, and functional operators into networks, enforcing constraints with losses, discovering relations from data, and 3D editing with guidance.

Invited Speakers

Leonidas Guibas

Leonidas Guibas

Stanford & Google

David Forsyth

David Forsyth

UIUC

Kostas Daniilidis

Kostas Daniilidis

Upenn

Richard Hao Zhang

Richard Hao Zhang

SFU

Maks Ovsjanikov

Maks Ovsjanikov

Ecole Polytechnique

Shenlong Wang

Shenlong Wang

UIUC

Rana Hanocka

Rana Hanocka

UChicago

Daniel Cremers

Daniel Cremers

TUM

Schedule

13:05 - 13:35 Speaker 1: Topic Title TBD Kostas Daniilidis
13:35 - 14:05 Speaker 2: Topic Title TBD Leonidas Guibas
14:05 - 14:35 Speaker 3: Topic Title TBD Richard Hao Zhang
14:35 - 15:05 Speaker 4: Topic Title TBD Rana Hanocka
15:05 - 15:25 Break
15:25 - 15:55 Speaker 5: Topic Title TBD Shenlong Wang
15:55 - 16:25 Speaker 6: Topic Title TBD Maks Ovsjanikov
16:25 - 16:55 Speaker 7: Topic Title TBD Daniel Cremers
16:55 - 17:25 Speaker 8: Topic Title TBD David Forsyth

Organizers

Qixing Huang

Qixing Huang

UT Austin

Hanwen Jiang

Hanwen Jiang

UT Austin

Congyue Deng

Congyue Deng

Stanford

Lin Gao

Lin Gao

UCAS

Biao Zhang

Biao Zhang

KAUST

Ruqi Huang

Ruqi Huang

Tsinghua SIGS

Anand Bhattad

Anand Bhattad

TTI-Chicago

Roni Sengupta

Roni Sengupta

UNC Chapel Hill

Despoina Paschalidou

Despoina Paschalidou

NVIDIA

Lingjie Liu

Lingjie Liu

UPenn