Causal Diffusion Autoencoders: Toward Counterfactual Generation via Diffusion Probabilistic Models

TL;DR

CausalDiffAE introduces a diffusion-based framework for causal representation learning that enables controllable counterfactual generation by disentangling causal variables and modeling their mechanisms, even with limited supervision.

Contribution

It proposes a novel causal encoding mechanism within diffusion models for high-level causal variable extraction and counterfactual generation, addressing interpretability and control in diffusion-based image synthesis.

Findings

Learns a disentangled causal latent space.

Capable of generating high-quality counterfactual images.

Effective even with limited label supervision.

Abstract

Diffusion probabilistic models (DPMs) have become the state-of-the-art in high-quality image generation. However, DPMs have an arbitrary noisy latent space with no interpretable or controllable semantics. Although there has been significant research effort to improve image sample quality, there is little work on representation-controlled generation using diffusion models. Specifically, causal modeling and controllable counterfactual generation using DPMs is an underexplored area. In this work, we propose CausalDiffAE, a diffusion-based causal representation learning framework to enable counterfactual generation according to a specified causal model. Our key idea is to use an encoder to extract high-level semantically meaningful causal variables from high-dimensional data and model stochastic variation using reverse diffusion. We propose a causal encoding mechanism that maps high-dimensional data to causally related latent factors and parameterize the causal mechanisms among latent factors using neural networks. To enforce the disentanglement of causal variables, we formulate a variational objective and leverage auxiliary label information in a prior to regularize the latent space. We propose a DDIM-based counterfactual generation procedure subject to do-interventions. Finally, to address the limited label supervision scenario, we also study the application of CausalDiffAE when a part of the training data is unlabeled, which also enables granular control over the strength of interventions in generating counterfactuals during inference. We empirically show that CausalDiffAE learns a disentangled latent space and is capable of generating high-quality counterfactual images.