Mine-Sample-Render: A Three-Stage Method for Diverse Generation

When diversity gets lost at the output

Modern generative models have achieved impressive quality — from text to images, from video to 3D assets. But there's a paradox: these same models consistently produce monotonous samples, even when explicitly asked for diversity. Competent text, but boring; quality images, but similar to the previous ones; correct user simulation, but repetitive.

Researchers Kostas Triaridis, Morris Sharp, and Nebojsa Jojic from Microsoft Research have proposed MINE-SAMPLE-RENDER (MSR) — an approach that turns the problem on its head. Instead of hoping for a model's built-in randomness mechanism to work, the authors propose an explicit, controlled procedure at test time. The idea is simple: take a corpus of real examples, extract a diversity map from it, then use that map to steer generation.

A three-stage pipeline

The first stage is mining. The method analyzes a corpus of real artifacts (weather reports, web designs, user descriptions) and extracts a categorical schema — the structure of attributes that vary in the data. In parallel, statistics for these attributes are computed, which will guide the subsequent stages.

The second stage is sampling. Based on the derived statistics, the method samples specific attribute configurations. These aren't arbitrary numbers, but a deliberate choice drawn from the distribution observed in real data. Attributes are chosen so that the resulting diversity reflects the diversity of the corpus.

The third stage is rendering. Here, the generative model serves purely as a conditional renderer. The model doesn't decide what to generate — that's already decided at the sampling stage. The model simply takes the attribute configuration and generates the corresponding output. This reduces the entropy of the process and improves reproducibility.

The method was tested across five different domains: weather report generation, user behavior simulation, web page design, text-to-image conversion, and agent-driven 3D asset creation. In all cases, MSR improved alignment with the corpus distribution and output diversity while preserving quality.

Context and related approaches

The problem of diversity in generative modeling is drawing active attention from researchers. On the Hugging Face platform, several closely related works have been discussed recently: • "Gacha Decoding: Eliciting Diverse Generations Through Instruction Following" (2026) • "Do Large Language Models Capture the Diversity in their Training Data?" (2026) • "Beyond Mode Collapse: Generating Diverse Synthetic Expert Conversations via Generative Flow Networks" (2026) • "DOGS: Design-Space Sampling for Prompt-Driven Logo Generation" (2026) • "Gap Detection: Mitigating Homogenization through Demand and Defensibility" (2026) • "A Ticket from Marginals to Joints: Coupled-Noise Distillation for One-Step Block Generation in Diffusion Language Models" (2026) • "DiVid: Diagnosing Dimension-Specific Diversity Collapse in Video Generation Models" (2026)

MSR stands out for its practicality: it's a test-time approach that doesn't require retraining the model. It can be applied to any existing generative model, making the method a universal tool for boosting output diversity.