[CVPR-25] Compass Control: Multi-Object Orientation Control for Text-to-Image Generation

Abstract

Compass Control enables precise and continuous object orientation control in text-to-image diffusion models. Given an input orientation angle along with a text prompt, our method can generate realistic scenes following the text prompts with object-centric orientation control.

Environment setup

The conda environment file is provided in environment.yaml. A list of explicit packages installed using pip is provided in explicit.txt. To setup the conda environment using the environment.yaml file, run

conda env create -f environment.yaml

Training

Dataset

You can download the dataset from this Google Drive link (around 1.5 GB). Place the zip file in the parent directory and run

unzip compass_dataset.zip

# activate the environment (need to set it up first) 
conda activate compass 

# run training! 
cd training_scripts/ 
chmod +x run.sh 
./run.sh 

Inference

Checkpoints

You can download the checkpoint from this Google Drive link (around 40 MB total). Place the folder __compass in ckpts/multiobject directory.

To visualize inference results, run training_scripts/infer_vis.ipynb in an interactive environment.