📄 中文摘要
一致性生成模型如Shortcut和MeanFlow通过目标感知设计在解决概率流常微分方程（PF-ODE）方面取得了显著成果。通常，这些方法在当前时间$t$旁引入目标时间$r$，以调节输出在局部多步导数（$r = t$）和全局少步积分（$r = 0$）之间的关系。然而，传统的“一输入一输出”范式强制划分训练预算，往往将大量资源（例如，MeanFlow中75%）单独分配给多步目标以确保稳定性。这种分离导致了一个权衡：为多步目标分配足够的样本会使少步生成训练不足，从而影响收敛性并限制可扩展性。
📄 English Summary
Duality Models: An Embarrassingly Simple One-step Generation Paradigm
Consistency-based generative models such as Shortcut and MeanFlow achieve remarkable results through a target-aware design for solving the Probability Flow ODE (PF-ODE). Typically, these methods introduce a target time $r$ alongside the current time $t$ to modulate outputs between a local multi-step derivative ($r = t$) and a global few-step integral ($r = 0$). However, the conventional 'one input, one output' paradigm enforces a partition of the training budget, often allocating a significant portion (e.g., 75% in MeanFlow) solely to the multi-step objective for stability. This separation imposes a trade-off: allocating sufficient samples to the multi-step objective leaves the few-step generation undertrained, which adversely affects convergence and limits scalability.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等