Close the performance gap with multistep generative models

Close the remaining performance gap between Priming Models and multi-step diffusion and flow models in generation quality and related evaluation metrics.

Background

Priming Models achieve strong one-step generation results across image generation and world modeling, but the paper acknowledges that multi-step diffusion and flow models remain a stronger reference in some settings. Matching or surpassing their remaining performance advantage is explicitly identified as an unresolved challenge.