---
license: cc-by-nc-nd-4.0
viewer: false
---
# Benchmarking Spatial Relationships in Text-to-Image Generation
*Tejas Gokhale, Hamid Palangi, Besmira Nushi, Vibhav Vineet, Eric Horvitz, Ece Kamar, Chitta Baral, Yezhou Yang*


- We introduce a large-scale challenge dataset SR<sub>2D</sub> that contains sentences describing two objects and the spatial relationship between them.
- We introduce a metric called VISOR (short for **V**erify**I**ng **S**patial **O**bject **R**elationships) to quantify spatial reasoning performance.  
- VISOR and SR<sub>2D</sub> can be used off-the-shelf with any text-to-image model.

## SR<sub>2D</sub> Dataset
Our dataset is hosted as  [here](https://huggingface.co/datasets/tgokhale/sr2d_visor).  This contains 
1. The text prompt dataset in `.json` format (`text_spatial_rel_phrases.json`)
2. Images generated using 7 models (GLIDE, CogView2, DALLE-mini, Stable Diffusion, GLIDE + Stable Diffusion + CDM, and Stable Diffusion v2.1) 

Alternatively, the text prompt dataset can also accessed from [`text_spatial_rel_phrases.json`](https://github.com/microsoft/VISOR/blob/main/text_spatial_rel_phrases.json). It contains all examples from the current version of the dataset (31680 text prompts) accompanied by the corresponding metadata.
This dataset can also be generated by running the script `python create_spatial_phrases.py`

## GitHub repository
The GitHub repository for [VISOR](https://github.com/microsoft/VISOR/) contains code for generating images with prompts from the SR<sub>2D</sub> dataset and evaluating the generated images using VISOR.


## References
Code for text-to-image generation:
1. GLIDE: https://github.com/openai/glide-text2im
2. DALLE-mini: https://github.com/borisdayma/dalle-mini
3. CogView2: https://github.com/THUDM/CogView2
4. Stable Diffusion: https://github.com/CompVis/stable-diffusion
5. Composable Diffusion Models: https://github.com/energy-based-model/Compositional-Visual-Generation-with-Composable-Diffusion-Models-PyTorch
6. OpenAI API for DALLE-2: https://openai.com/api/

## Citation
If you find SR<sub>2D</sub> or VISOR useful in your research, please use the following citation:
```
@article{gokhale2022benchmarking,
  title={Benchmarking Spatial Relationships in Text-to-Image Generation},
  author={Gokhale, Tejas and Palangi, Hamid and Nushi, Besmira and Vineet, Vibhav and Horvitz, Eric and Kamar, Ece and Baral, Chitta and Yang, Yezhou},
  journal={arXiv preprint arXiv:2212.10015},
  year={2022}
}
```