simple-rectified-flow

A simple implementation for Rectified Flow. Some of the framework references COMPSCI 180 FA24, UC Berkeley.

Requirements

Please refer to & install by

conda create -n rf python=3.12
conda activate rf
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124 # tailor the CUDA version to your needs
pip install -r requirements.txt

Usage

Before training

python prepare.py

Training of Rectified Flow by MNIST

Time-conditional Rectified Flow (Unconditional)

python train_rf_timecond.py

Class-conditional Rectified Flow

python train_rf_classcond.py