Publikation
Beyond Overcorrection: Evaluating Diversity in T2I Models with DivBench
Felix Friedrich; Thiemo Ganesha Welsch; Manuel Brack; Patrick Schramowski; Kristian Kersting
In: Computing Research Repository eprint Journal (CoRR), Vol. abs/2507.03015, Pages 1-7, Computing Research Repository, 2025.
Zusammenfassung
Current diversification strategies for text-to-image (T2I)
models often ignore contextual appropriateness, leading to
over-diversification where demographic attributes are mod-
ified even when explicitly specified in prompts. This paper
introduces DIVBENCH, a benchmark and evaluation frame-
work for measuring both under- and over-diversification in
T2I generation. Through systematic evaluation of state-of-
the-art T2I models, we find that while most models exhibit
limited diversity, many diversification approaches overcor-
rect by inappropriately altering contextually-specified at-
tributes. We demonstrate that context-aware methods, par-
ticularly LLM-guided FairDiffusion and prompt rewriting,
can already effectively address under-diversity while avoid-
ing over-diversification, achieving a better balance between
representation and semantic fidelity.