Seonglae ChoTo systematically train the first large-scale and stable "Continuous-time CM," we integrate EDM's parameterization with the Flow Matching perspective (TrigFlow) and systematically address the instability issues in CM training. CT-CM's formulation is elegantly expressed through TrigFlow.
TrigFlow
Simplifying, Stabilizing and Scaling Continuous-time Consistency...
Consistency models (CMs) are a powerful class of diffusion-based generative models optimized for fast sampling. Most existing CMs are trained using discretized timesteps, which introduce additional...
https://openreview.net/forum?id=LyJi5ugyJx