The preprint "Are We Overthinking One-Step Generation?" (arXiv:2610.20331) begins with a familiar observation: diffusion and flow matching models are now the dominant frameworks for generative modeling in continuous domains, but they typically require multiple network evaluations at inference. This computational burden has motivated a large body of research aimed at achieving one-step generation.
The paper's title raises the question of whether that research direction has become needlessly complicated. However, the abstract excerpt provided here only lays out the motivation; it does not include the authors' argument or findings. The full paper would be needed to see what they propose in response.