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

Key points

• Modern generative models produce high-quality individual outputs but struggle to sample diversely, even when explicitly instructed to do so.1
• Our pipeline (i) mines an interpretable categorical schema from a corpus of real artifacts and supplies statistics for sampling; (ii) samples attribute configurations from the corpus-derived statistics; (iii) uses a generative model purely as a conditional renderer of a sampled configuration.1
• Across weather report generation, user simulation, webpage design, text-to-image generation, and agent-generated 3D assets, we evaluate the diversity and output quality of MSR-guided generation and show that MSR improves alignment with corpus distributions and diversity of generated outputs.1

Sources (2)

• [1]Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 03:40 AM

  We introduce MINE-SAMPLE-RENDER (MSR), a test-time approach that externalizes the diversity into an explicit, corpus-derived sampling procedure.

  Modern generative models produce high-quality individual outputs but struggle to sample diversely, even when explicitly instructed to do so.

• [2]Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models
  Hugging Face Daily Papers · Oct 15, 12:00 AM · same content

Extractive summary: sentences quoted from the sources.