Explorative modeling: Train on the best of K guesses | Hacker News

A Hacker News discussion highlighted a paper that adapts winner-take-all training to multimodal diffusion and flow models.

The method learns from only the best of K candidate guesses, aiming to reduce mode collapse in multimodal outputs.

Critics pointed to extra training forward passes and sampling that treats all modes equally, which may hurt efficiency and quality.

The commenter argued the approach could be improved with a different implementation, but may be less useful for highly conditioned image generation and for LLMs.