SharpEuler: Training-Free Sampler Optimizes Flow Matching Steps
Researchers have unveiled SharpEuler, an innovative sampler that eliminates the need for training in flow matching models. It improves the integration process by focusing on areas with the most significant changes in the learned velocity field. The method assesses sharpness using finite-difference along calibration paths and then converts this into a non-uniform timestep grid via a quantile transformation. For inference, it sticks to standard Euler integration, keeping the model evaluations consistent with a uniform schedule. This approach is grounded in three fundamental concepts: numerical acceleration, solver-aware profiling, and budget-adaptive scheduling. You can find this research on arXiv under ID 2605.11547.
Key facts
- SharpEuler is a training-free sampler for flow matching models.
- It profiles a pretrained model offline by estimating velocity field sharpness.
- Sharpness is estimated via finite-difference along calibration trajectories.
- A quantile transform converts the sharpness profile into a timestep grid.
- At test time, it uses Euler integration with the same number of steps as uniform.
- The method is justified by three principles: numerical acceleration, solver-aware profiling, budget-adaptive scheduling.
- Published on arXiv with ID 2605.11547.
- The approach aims to improve fast generation under a fixed evaluation budget.
Entities
Institutions
- arXiv