DOS: Directional Object Separation in Text Embeddings (AAAI 2026)

This repository contains the official implementation of the paper: DOS: Directional Object Separation in Text Embeddings for Multi-Object Image Generation

Prerequisites

The code was tested on a RTX3090 but should work on other cards with at least 24GB VRAM.

conda env create --file environment.yaml
conda activate dos
pip install -e .

Note: Please input the valid hugging face token in ./configs/envs.py before running diffusion models.

How to test

Please refer to the following notebooks:

• SDXL: ./notebooks/test_dos_sdxl.ipynb
• SD3.5: ./notebooks/test_dos_sd3.5.ipynb

How to run benchmark

DATASET="similar_shapes"

# SDXL
python run_benchmark.py \
--device cuda:0 \
--output_path outputs/performance_comparison/${DATASET}/sdxl \
--dataset ${DATASET} \
--method sdxl \
--seed_range 1 5

# SDXL with DOS
python run_benchmark.py \
--device cuda:0 \
--output_path outputs/performance_comparison/${DATASET}/sdxl_with_dos \
--dataset ${DATASET} \
--method sdxl_with_dos \
--lambda_sep 1.0 \
--seed_range 1 5

How to evaluate

We measure gpt-4o-mini-based Sucess Rate (SR) and Mixture Rate (MR) for the generated images.

python evaluate_with_vlm.py \
--folder outputs/performance_comparison/similar_shapes/sdxl \
--model openai/gpt-4o-mini \
--api_key {your_open_router_api_key}

For more details, please refer to the script evaluate_with_vlm.py.

Citation

@article{byun2025directional,
  title={DOS: Directional Object Separation in Text Embeddings for Multi-Object Image Generation},
  author={Byun, Dongnam and Park, Jungwon and Ko, Jumgmin and Choi, Changin and Rhee, Wonjong},
  journal={arXiv preprint arXiv:2510.14376},
  year={2025}
}