MSR: Test-Time Method Externalizes Diversity to Corpus-Derived Sampling for Generative Models
Kostas Triaridis · hf · 2026-10-17
Researchers from Stony Brook University and Microsoft propose Mine, Sample, Render (MSR), a test-time alignment approach targeting generative models' poor sampling diversity despite high individual output quality.
The pipeline externalizes diversity into an explicit, corpus-derived sampling procedure:
- Mine: extract an interpretable categorical schema and statistics from a corpus of real artifacts;
- Sample: draw attribute configurations from the corpus-derived statistics;
- Render: use the generative model purely as a conditional renderer of the sampled configuration.
Evaluations across weather report generation, user simulation, webpage design, text-to-image generation, and agent-generated 3D assets show MSR improves alignment with corpus distributions and output diversity.
More from Multimodal
- Community turns to new MOSS Image model as Flux-3 Klein disappoints — Cultural-Umpire9061 · 2026-10-17
- Miles-Diffusion adds RL training for FLUX 3 Image Klein, runs on a single node — BanghuaZ · 2026-10-17
- AI video demo 'Omni 1'1 flash' wows Reddit with near-photoreal cinematic footage — Dense-Instruction210 · 2026-10-17
- Redditor recreates Austin Powers scene with AI video, 'Mini Me' likeness impresses — Alternative_Can_8663 · 2026-10-17
- N8Vision: open-source image browser for managing 50k+ AI-generated images — Consistent_Help2177 · 2026-10-17
- Dev trains a 'microdrama model': 'my wallet is gonna thank me' — Kyrannio · 2026-10-17