Model-Aware Schedules Improve Generation via Fiberwise Optimal Transport
The paper reports model-aware diffusion and flow-matching schedules that beat strong schedule baselines, including a 38.6% relative FID drop on CIFAR-10 flow matching at 16 function evaluations.
The authors build schedules by adding a fiberwise prediction-risk term to the kinetic action used in optimal-transport-motivated schedule design. That risk can be estimated from an early baseline checkpoint, then used to choose a closed-form time allocation along a fixed coefficient curve. Across DDPMs and flow matching settings, the resulting schedules consistently improved generation in the paper’s evaluations. The authors also report that normalized risk profiles and schedule deformations align across independently trained models, enough that a frozen analytic template keeps most of the gain. ArXiv · AI/CL/LG's note
The authors build schedules by adding a fiberwise prediction-risk term to the kinetic action used in optimal-transport-motivated schedule design. That risk can be estimated from an early baseline checkpoint, then used to choose a closed-form time allocation along a fixed coefficient curve. Across DDPMs and flow matching settings, the resulting schedules consistently improved generation in the paper’s evaluations. The authors also report that normalized risk profiles and schedule deformations align across independently trained models, enough that a frozen analytic template keeps most of the gain. ArXiv · AI/CL/LG's note
score 5