Causal Inference and Generation

Introduction

Deep generative models have achieved remarkable success in synthesizing realistic images and other high-dimensional data. However, standard generative models learn statistical correlations rather than the underlying causal mechanisms governing data generation. As a result, they cannot reliably answer interventional or counterfactual queries: what would this image look like if a specific attribute were changed, while everything else remained the same? Equipping generative models with causal reasoning capabilities is essential for applications that require principled manipulation of data under causal constraints, including domain generalization, data augmentation, fairness, and explainability.

Counterfactual generation requires executing Pearl's three-step procedure — abduction, action, and prediction. While this procedure is well-defined theoretically, realizing it faithfully in high-dimensional image generative models introduces several challenges. A central challenge is accurately preserving the non-observed exogenous factors — such as identity, texture, and lighting — that should remain unchanged under a causal intervention. These factors must be precisely abducted from the observed image and disentangled from the intervened attributes, so that the generated counterfactual reflects only the intended causal change. Furthermore, most methods rely on statistical correlations in training data rather than true causal structure, causing spurious co-changes in unrelated attributes under intervention.

Our recent works address these challenges by developing methods that augment generative models with causal inference capabilities, enabling faithful interventions and counterfactual generation that preserves non-observed factors, yielding faithful counterfactuals, and satisfy Pearl's axiomatic criteria of composition, effectiveness, and reversibility. The resulting methods support downstream applications including out-of-distribution generalization, counterfactual data augmentation, causal image editing, and fairness.

Recent Works

Faithful Counterfactual Image Generation via Noise Abduction

This line of research addresses the problem of faithfully abducting and preserving exogenous noise variables in an SCM during counterfactual image generation. The non-observed exogenous factors that encode identity, texture, and other instance-specific attributes follow complex, multimodal posteriors that are difficult to capture accurately. We propose methods that learn latent representations compatible with this complexity, ensuring that the abducted noise faithfully preserves non-intervened, and non-observed factors so that the generated counterfactual reflects only the intended causal change.

Faithful Counterfactual Image Generation via Disentangled Noise Abduction

This research addresses the problem of faithfully abducting and preserving exogenous noise variables during counterfactual image generation under a structural causal model. A key challenge is that the abducted exogenous noise must be precisely disentangled from the intervened attributes, ensuring that non-observed factors such as identity and texture are retained without leakage from attribute information. We propose a method that achieves this via adversarial refinement, learning to separate attribute information from the abducted noise while preserving non-intervened identity information. The novelties of this work lie in 1) proposing an adversarial disentanglement objective that removes attribute information from the abducted noise while preserving identity. 2) demonstrating improvements over state-of-the-art methods on Pearl's axiomatic counterfactual criteria of composition, effectiveness, and reversibility.

Hierarchical Noise Abduction for Identity-Preserving Counterfactual Generation

This research extends noise abduction to a hierarchical latent structure, addressing the challenge that the exogenous noise is semantically heterogeneous and operates at distinct abstraction levels. We show that the complexity of the posterior over non-observed factors grows multiplicatively in the number of unmodeled factors, and propose a method that decomposes exogenous noise across a hierarchy of abstraction levels, reducing this representational cost from multiplicative to additive. By routing targeted refinement according to a per-level attribute leakage spectrum, the method preserves non-observed factors at each abstraction level more faithfully than flat abduction approaches.