Benchmarking Spatial Relationships in Text-to-Image Generation

You can download the report from here.

For further questions, please reach out to Tejas Gokhale (gokhale@umbc.edu) and Hamid Palangi (hpalangi@microsoft.com)

• We introduce a large-scale challenge dataset SR_2D that contains sentences describing two objects and the spatial relationship between them.
• We introduce a metric called VISOR (short for VerifyIng Spatial Object Relationships) to quantify spatial reasoning performance.
• VISOR and SR_2D can be used off-the-shelf with any text-to-image model.

SR_2D Dataset

Our dataset is hosted as a Huggingface Dataset here. 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. 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

Generating Images

In our experiments, we generated images for each text prompt in SR_2d for 5 models: GLIDE, DALLE-mini, CogView2, DALLE-v2, Stable Diffusion (SD) as well as Composable Diffusion Models (CDM) with GLIDE (GLIDE + CDM) and SD (SD + CDM). See acknowledgments below for the corresponding codebases / APIs.

In this repository, we provide a few output images in the images directory, as well as a webpage or viz.html for visualizing them.

Scripts for Evaluation

Formatting for Object Detection results

In objdet_results we provide .json files that contain object detection results using OWL-ViT for images generated by eight state-of-the-art diffusion models (GLIDE, GLIDE+CDM, DALLE-mini, CogView2, DALLE-v2, Stable Diffusion, SD+CDM, and SD-2.1). The format of these results file is simple -- each file contains a dictionary with keys being the filenames of images and values being the object detection results for that image. These object detection results are themselves dictionaries, which contain a list of classes, corresponding centroids of bounding boxes, recall (number of detected objects), sra (recall divided by total number of expected objects) and the text prompt.

An example is shown below:

{
"1103_0": {
    "classes": [
      "kite"
    ],
    "centroid": [
      [
        170.64540100097656,
        135.6916046142578
      ]
    ],
    "recall": 1,
    "sra": 0,
    "text": "a kite to the right of a bicycle"
  },
  "11036_1": {
    "classes": [
      "cat"
    ],
    "centroid": [
      [
        170.22688674926758,
        108.26640701293945
      ]
    ],
    "recall": 1,
    "sra": 0,
    "text": "a cat below a bottle"
  }
}

If you want to use your own results, please convert them to the above format and name the file results_<model_name>_<dataset_file_name>_<object_detector_name>_<threshold>.json. For example: results_dallev2_text_spatial_rel_phrases_owlvit_0.1.json

Running VISOR

Once the object detection results are setup as described above, run the visor evaluation code:

python visor.py --text_json <path_to_SR2D_data> 

We also provide the functionality of running the evaluation for specific models as comma-separated arguments:

python visor.py --text_json <path_to_SR2D_data> --mo glide,dallev2,stable-diffusion

Notebook for Evaluation

The notebook text2image_realtime_evaluate.ipynb offers a walk-through for using VISOR to evaluate the generated images and obtaining the VISOR scores for each model. This notebook also includes the pipeline for using OWL-ViT object detection.

Results from the Paper

Benchmarking results are shown below:

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_2D 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}
}