RAIGen: Rare Attribute Identification in Text-to-Image Generative Models

Text-to-image diffusion models achieve impressive generation quality but inherit and amplify training-data biases, skewing coverage of semantic attributes. Prior work addresses this in two ways. Closed-set approaches mitigate biases in predefined fairness categories (e.g., gender, race), assuming socially salient minority attributes are known a priori. Open-set approaches frame the task as bias identification, highlighting majority attributes that dominate outputs. Both overlook a complementary task: uncovering rare or minority features underrepresented in the data distribution (social, cultural, or stylistic) yet still encoded in model representations. We introduce RAIGen, the first framework, to our knowledge, for label-free rare-attribute discovery in diffusion models, requiring no predefined minority categories. RAIGen leverages Matryoshka Sparse Autoencoders and a novel minority metric combining neuron activation frequency with semantic distinctiveness to identify interpretable neurons whose top-activating images reveal underrepresented attributes. Experiments show RAIGen discovers attributes beyond fixed fairness categories in Stable Diffusion, scales to larger models such as SDXL, supports systematic auditing across architectures, and enables targeted amplification of rare attributes during generation. The project page is available at https://vssilpa.github.io/RAIGen_webpage/.

AI tools that turn text into images learn from pictures scraped from the internet. Because those collections are lopsided, the tools tend to overproduce whatever was common in the data and quietly neglect everything that was rare such as unusual cultural details, styles, or features that appeared only occasionally. Earlier research has tackled this from two angles: either fixing bias in categories we already know to worry about, such as gender or race, or pointing out which features dominate the outputs. But nobody had asked the opposite question: what rare things has the model actually learned, even though they almost never show up? We built RAIGen, the first method that uncovers these hidden, underrepresented features on its own, without being told in advance what to look for. It inspects the model's internal wiring, singling out the components that fire only for rare and distinctive content, then shows the images that trigger them so a person can see what was found. This lets researchers audit image generators more thoroughly, reveal blind spots, and work toward systems that represent the world more fully, and the approach scales to larger models too.