Modern generative models can produce high-quality individual outputs, but they often fail to sample diversely, even when explicitly prompted to do so. This is the problem addressed by a new arXiv paper introducing MINE-SAMPLE-RENDER (MSR).
MSR is a test-time approach that externalizes part of the generation process, according to the abstract. The idea is to shift some aspect of sampling outside the model itself, though the released abstract is cut off before detailing the full mechanism or results.
The paper highlights a known tension between output quality and diversity, and proposes a method that operates at inference time rather than requiring retraining. As the abstract stands, it does not yet provide experimental evidence or a complete description of the externalization step.