Unified Latents for Fair and Interpretable Generation: Latent-Space Debiasing and Attribute Control

TL;DR: Let’s make UL not just powerful, but fair and transparent—by integrating a fairness-promoting regularizer and explicit attribute disentanglement into its latent space. This could involve learning latent directions for sensitive attributes without predefined labels, enabling post-hoc debiasing and interpretable editing.

Research Question: Can latent representations learned by UL be structured to both mitigate bias (across sensitive or underrepresented groups) and support direct, interpretable control over semantic attributes?

Hypothesis: Integrating fairness-promoting regularizers (as in DDM) and disentanglement strategies (as in Dynamic Gaussian Anchoring) into UL’s latent space will yield generative models that produce fairer outputs and empower users to control or audit latent attributes.

Experiment Plan: Augment the UL objective with a fairness-promoting regularizer that balances latent representations across groups, without explicit attribute supervision. Integrate attribute-separating inductive biases (e.g., Dynamic Gaussian Anchoring or geometric regularizers). Train on datasets known for demographic imbalance (e.g., FairFace, CelebA). Measure bias (demographic parity, equalized odds) and attribute disentanglement (mutual information, interpretability of latent directions) in generated samples. Evaluate both quality (FID, PSNR) and fairness/interactivity compared to baseline latent diffusion and debiasing models.

References:

• Huang, L.-C., Tsao, C. C., Su, F.-Y., & Chiang, J.-H. (2025). Debiasing Diffusion Model: Enhancing Fairness through Latent Representation Learning in Stable Diffusion Model. arXiv.org.
• Jun, Y., Park, J., Choo, K., Choi, T. E., & Hwang, S. J. (2024). Disentangling Disentangled Representations: Towards Improved Latent Units via Diffusion Models. IEEE Workshop/Winter Conference on Applications of Computer Vision.