Counterfacts Without an Unnecessary Black Box: Diffusion Itself Explains Its Decisions

Counterfactual explanations in computer vision answer a simple question: What is the smallest change to an image that would overturn the model’s prediction? In medicine, quality control, and other safety-critical fields, this is not just an academic exercise—it’s a way to understand why a model made a mistake or took a risky decision.

Existing approaches based on diffusion models can generate realistic modifications, but they run into an architectural limitation: they require an external classifier that must perform reliably on noisy intermediate images. Essentially, to explain one model, you add a second one—and must trust it in conditions where it is known to be unstable. Two black boxes instead of one.

C-VCE solves this problem differently. The classifier is embedded directly into the generative model via a concept bottleneck layer—a layer that operates on human-readable features (concepts) rather than pixels. Counterfactual adjustments are guided by these concepts from within, without a separate supervisor model.

The user can enable and disable semantic concepts directly during sampling. The model makes minimal adjustments only to relevant regions of the image, leaving the rest untouched and accounting for correlations between features. To control the extent of the edits, we added a probabilistic regularizer that balances between “flip the prediction” and “stay close to the original,” plus a gradient mask that limits changes to the most significant regions.

On the CelebA C-VCE benchmark, our method matches or outperforms baseline methods in terms of flip rate (the proportion of successfully flipped predictions), while the counterfacts are visually closer to the original image and less distorted. This is to be expected: when the classifier does not have to deal with noise separately from the generator, the edits turn out more accurate.

The main conclusion goes beyond this specific method. Uncovering and controlling the internal layer of concepts within a generative model is a promising path toward making diffusion systems more understandable and safer. Instead of imposing explainability from the outside via additional models, it can be built into the architecture. This changes the very architecture of trust: you trust a single model with a transparent layer, rather than a stack of two opaque ones.

One question that remains open is to what extent the concept-bottleneck approach scales to domains beyond faces and attributes like CelebA—for example, to medical images, where concepts are less obvious and their interrelationships are more complex.

Source: cs.AI updates on arXiv.org