Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models

Mine, Sample, Render. Test-Time Diversity Alignment Of Generative Models. By Kostas Triaridis and two others, from Stony Brook University and Microsoft. Published on the fifteenth of October. No code repository is linked to it.

Modern generative models produce high-quality individual outputs, but they struggle to sample diversely. That holds even when they are explicitly told to be diverse. Ask for variety and the model, left to itself, still does not deliver it.

MINE-SAMPLE-RENDER, or MSR, is a test-time approach that takes diversity out of the model's hands. It moves the diversity into an explicit sampling procedure, derived from a corpus of real artifacts. The model is no longer asked to be varied. It is asked to draw what it is told.

The pipeline has three steps, and they are the three words of the name. First, mine: an interpretable categorical schema is extracted from a corpus of real artifacts, together with statistics for sampling. Second, sample: attribute configurations are drawn from those corpus-derived statistics. Third, render: the generative model is used purely as a conditional renderer, turning one sampled configuration into an output. The variety comes from the sampling step, and the model supplies the quality.

The evaluation is broad in domain. It covers weather report generation, user simulation, webpage design, text-to-image generation, and 3D assets generated by agents. Across these, the authors measure both the diversity and the quality of outputs guided by MSR, and report that it improves alignment with the corpus distributions and the diversity of what gets generated.

The move is a division of labour. Rather than coaxing a model into diversity it does not naturally produce, MSR hands the choice of what to make to an interpretable, corpus-derived procedure, and leaves the model the part it already does well: rendering one thing at high quality.