CAR Guidance — Image Experiment

Controlling Rectified Flow on CelebA-HQ

The pre-trained generative model can be downloaded from Rectified Flow CelebA-HQ Just put it in ./

Environment Setup

The project runs in a conda environment named car_guidance_image (Python 3.10, PyTorch CUDA build). The provided environment.yml is a full export of the author's environment; use it to reproduce the environment exactly:

conda env create -f environment.yml
conda activate car_guidance_image

Note: if you hit ImportError: ...libtorch_cpu.so: undefined symbol: iJIT_NotifyEvent, it's a missing runtime libittnotify.so in some environments. The fix is to build a minimal stub library and preload it. For example:

gcc -shared -fPIC -Wl,-soname,libittnotify.so -x c -o "${CONDA_PREFIX}/lib/libittnotify.so" - <<'EOF'
int iJIT_NotifyEvent(int t, void *d) { (void)t; (void)d; return 0; }
int iJIT_IsProfilingActive(void) { return 0; }
unsigned int iJIT_GetNewMethodID(void) { static unsigned int id = 1; return id++; }
EOF
export LD_PRELOAD="${CONDA_PREFIX}/lib/libittnotify.so${LD_PRELOAD:+:${LD_PRELOAD}}"

Run

We provide a demo image ./demo/celeba.jpg for running our model.

python main_data.py

Dataset

The full Celeba-hq-1024 dataset can be downloaded from kaggle celeba-hq dataset