Minimizing Trajectory Curvature of ODE-
based Generative Models
Sangyun Lee, Beomsu Kim, Jong Chul Ye
https://cvpr2022-tutorial-diffusion-models.github.io/
Probability Flow ODE
Truncation error
Probability Flow ODE
Accurate!
Straight generative path
intersection!
intersection!
Coupling Degree of intersection
Training objective (q)
Training objective (joint)
Upper bound
Coupling Degree of intersection
Training objective (q)
Training objective (joint)
Upper bound
Coupling Degree of intersection
Training objective (q)
Training objective (joint)
Upper bound
Coupling Degree of intersection
Training objective (q)
Training objective (joint)
Upper bound
Coupling Degree of intersection
Training objective (q)
Training objective (joint)
Upper bound
Coupling Degree of intersection
Training objective (q)
Training objective (joint)
Upper bound
NFEs
=
4
NFEs
=
128
Independent Ours
Low-curvature generative ODEs make less distillation error.
Conclusion
● Trajectory curvature is directly related to sampling efficiency of diffusion
models.
● The trajectory curvature can be reduced by learning dependence between
data and noise.
● Our method reduces sampling costs and improves distillation performance
of diffusion models.
Check out github.com/sangyun884/fast-ode for code and updates.

Minimizing Trajectory Curvature of ODE-based Generative Models.pdf