Search papers, labs, and topics across Lattice.
The paper introduces CompDiff, a hierarchical compositional diffusion framework, to address the imbalanced generator problem in medical image generation, where generative models produce lower quality images for rare demographic subgroups. CompDiff uses a Hierarchical Conditioner Network (HCN) to decompose demographic conditioning into a demographic token concatenated with CLIP embeddings, promoting parameter sharing and compositional generalization. Experiments on chest X-rays and fundus images demonstrate that CompDiff improves image quality, subgroup equity, and zero-shot intersectional generalization compared to fine-tuning and FairDiffusion, leading to improved downstream classifier performance and reduced bias.
A new diffusion architecture that explicitly disentangles demographic factors allows for generating higher-quality medical images for underrepresented groups and novel demographic intersections, outperforming standard fine-tuning and FairDiffusion.
Generative models are increasingly used to augment medical imaging datasets for fairer AI. Yet a key assumption often goes unexamined: that generators themselves produce equally high-quality images across demographic groups. Models trained on imbalanced data can inherit these imbalances, yielding degraded synthesis quality for rare subgroups and struggling with demographic intersections absent from training. We refer to this as the imbalanced generator problem. Existing remedies such as loss reweighting operate at the optimization level and provide limited benefit when training signal is scarce or absent for certain combinations. We propose CompDiff, a hierarchical compositional diffusion framework that addresses this problem at the representation level. A dedicated Hierarchical Conditioner Network (HCN) decomposes demographic conditioning, producing a demographic token concatenated with CLIP embeddings as cross-attention context. This structured factorization encourages parameter sharing across subgroups and supports compositional generalization to rare or unseen demographic intersections. Experiments on chest X-rays (MIMIC-CXR) and fundus images (FairGenMed) show that CompDiff compares favorably against both standard fine-tuning and FairDiffusion across image quality (FID: 64.3 vs. 75.1), subgroup equity (ES-FID), and zero-shot intersectional generalization (up to 21% FID improvement on held-out intersections). Downstream classifiers trained on CompDiff-generated data also show improved AUROC and reduced demographic bias, suggesting that architectural design of demographic conditioning is an important and underexplored factor in fair medical image generation. Code is available at https://anonymous.4open.science/r/CompDiff-6FE6.
