Fair Image Generation of Minority Groups

Image classifiers inherit the biases of their training data. On CelebA, a well-known correlation is that women are far more likely to be smiling than men — so a naive classifier learns to lean on gender as a shortcut for smiling. The research question I set out to test:

Can a structural causal model (SCM), used to disentangle a protected attribute in the latent space of a bidirectional generative model, generate new images of the under-represented group well enough to correct that bias downstream?

I built on DEAR (weakly-supervised disentangled generative causal representation learning), which embeds an SCM over the latent variables of a bidirectional GAN so that latent factors map to semantically meaningful, causally-structured attributes. The pipeline:

1. Inject a known bias into the CelebA training distribution (the smiling/gender correlation) so I had ground truth to measure against.
2. Learn a disentangled, causal latent representation so I could intervene on one attribute while holding others fixed.
3. Generate counterfactual minority-group samples — e.g. smiling men — by intervening in latent space rather than hunting for rare real examples.
4. Re-train the supervised classifier on the augmented set and measure whether the gender shortcut weakened.

This was my deepest dive into the seam between causal inference and deep generative modelling, and it taught me how hard honest disentanglement is. Reading DEAR closely — then bending its assumptions to a fairness setting — was a different kind of work from my from-scratch reimplementations: here the contribution was the experimental design and the causal framing, not the architecture. It's also the project that most shaped how I think about evaluation: a generative model that produces pretty samples can still be useless if the latent factors aren't actually doing what you claim.

CelebA bias is a clean, almost toy setting; real-world protected attributes are messier and rarely binary. Disentanglement quality is also notoriously sensitive and hard to measure, so claims here should be read as directional evidence under a controlled bias, not a deployable de-biasing recipe. The full method, metrics, and figures are in the linked report.