Learning generalizable visual representations with causal diffusion model for controllable editing

Shanshan Huang

doi:10.1016/j.patcog.2026.113162

Learning generalizable visual representations with causal diffusion model for controllable editing

2026-01-29

0

PRE

AI

DOI：10.1016/j.patcog.2026.113162

Original

Original request for help

Save

Abstract

En 中文

• A diffusion model-based causal representation learning method, CausalDiffuser, is proposed. • The representations are divided into causal and stochastic representations to ensure the image reconstruction quality. • Structural causal model serves as priors, capturing the causal relationships among the underlying generative factors (i.e., visual attributes). • A composite loss is designed to enhance the causal representation learning capability by introducing supervisory information. • The effectiveness of CausalDiffuser is validated on synthetic and real-world datasets.

Keywords:

causal representation learning

diffusion model

structural causal model

controllable editing

generalizable visual representations

Learning generalizable visual representations with causal diffusion model for controllable editing

Abstract

Journal

Organization