Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models
MSR puts diversity in a corpus-based sampling step, leaving the model to render the chosen configuration.
The method mines an interpretable set of categories and sampling statistics from real artifacts. It then samples attribute combinations and asks a generative model to produce outputs matching them. Across weather reports, user simulation, webpage design, text-to-image generation and agent-generated 3D assets, the authors report greater output diversity and closer alignment with corpus distributions.
HF Daily Papers' note
The method mines an interpretable set of categories and sampling statistics from real artifacts. It then samples attribute combinations and asks a generative model to produce outputs matching them. Across weather reports, user simulation, webpage design, text-to-image generation and agent-generated 3D assets, the authors report greater output diversity and closer alignment with corpus distributions.
HF Daily Papers' note
score 5