[Project Page] [arXiv] [ComfyUI node]

Gaoyang Zhang, Bingtao Fu, Qingnan Fan, Qi Zhang, Runxing Liu, Hong Gu, Huaqi Zhang, Xinguo Liu
ICCV 2025

CoMPaSS enhances the spatial understanding of existing text-to-image diffusion models, enabling them to generate images that faithfully reflect spatial configurations specified in the text prompt.

We manage our python environment with uv, and provide a convenient script for setting up the environment at setup_env.sh. Running this script will create a subdirectory .venv/ in the project root. To enable it, run source .venv/bin/activate after the environment is set up:

# install requirements into .venv/
bash ./setup_env.sh
# activate the environment
source .venv/bin/activate

We recommend trying out the FLUX.1-dev LoRA trained via CoMPaSS. Please refer to the custom node's repository to get started.

We provide the reference weights used to report all metrics in our paper on Hugging Face 🤗. We recommend trying out the FLUX.1-dev weights as it is a Rank-16 LoRA which is only 50MB in size.

We provide full instructions for replicating the SCOP dataset (28,028 object pairs among 15,426 images) in the SCOP directory. Check out its README to get started.

We provide both training and inference instructions for using our TENOR module in the TENOR directory. MMDiT-based models (e.g., FLUX.1-dev) and UNet-based models (e.g., SD1.5) are both supported. Check out their respective instructions to get started:

• Instructions for FLUX.1-dev
• Instructions for SD1.4, SD1.5, and SD2.1

@inproceedings{zhang2025compass,
  title={CoMPaSS: Enhancing Spatial Understanding in Text-to-Image Diffusion Models},
  author={Zhang, Gaoyang and Fu, Bingtao and Fan, Qingnan and Zhang, Qi and Liu, Runxing and Gu, Hong and Zhang, Huaqi and Liu, Xinguo},
  booktitle={ICCV},
  year={2025}
}