willisma / SiT

Official PyTorch Implementation of "SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers"
https://scalable-interpolant.github.io/
MIT License
662 stars 35 forks source link

Maybe inconsistent behavior of the method compute_drift() in VPCPlan and ICPlan ? #17

Closed zhengyunqq closed 5 months ago

zhengyunqq commented 5 months ago

In the ICPlan, compute_drift() returns (-drift, diffusion).

However, in VPCPlan, I think compute_drift() returns (drift, diffusion). drift = d_alpha_alpha_ratio x, for VPCPlan, d_alpha_alpha_ratio= -0.5 \beta but compute_drift() in VPCPlan returns -0.5\beta instead of 0.5\beta

Their signs on drift are incosistent.