PyTorch code for the EMNLP 2021 paper "Integrating Visuospatial, Linguistic and Commonsense Structure into Story Visualization". See the arxiv paper here.

I have discovered an error in the FID evaluation of our model outputs, thanks to the participants in Issue #6. The prior FID calculations were made using model outputs as reference (due to a hard-coded directory path :-( ), hence the low values. I apologize for the inconvenience it has caused to the users of both the repositories. I made fresh calculations using the correct reference and pretrained checkpoints, see the updated results below. We will update the arxiv paper with these results in the next few days.

FID Scores (↓ is better)

This code has been tested on torch==1.11.0.dev20211014 (nightly) and torchvision==0.12.0.dev20211014 (nightly)

Download the PororoSV dataset and associated files from here and save it as ./data. Download GloVe embeddings (glove.840B.300D) from here. The default location of the embeddings is ./data/ (see ./dcsgan/miscc/config.py).

Download the data files for Flintstones here. Download the full Flintstones data here.

To install the Berkeley Neural Parser with SpaCy:

pip install benepar

To extract parses for PororoSV:

python parse.py --dataset pororo --data_dir <path-to-data-directory>

We use the Dense Captioning Model implementation available here. Download the pretrained model as outlined in their repository. To extract dense captions for PororoSV:
python describe_pororosv.py --config_json <path-to-config> --lut_path <path-to-VG-regions-dict-lite.pkl> --model_checkpoint <path-to-model-checkpoint> --img_path <path-to-data-directory> --box_per_img 10 --batch_size 1

To train VLC-StoryGAN for PororoSV:
python train_gan.py --cfg ./cfg/pororo_s1_vlc.yml --data_dir <path-to-data-directory> --dataset pororo\

Unless specified, the default output root directory for all model checkpoints is ./out/

Please see here for evaluation models for character classification-based scores, BLEU2/3 and R-Precision.

To evaluate Frechet Inception Distance (FID):
python eval_vfid --img_ref_dir <path-to-image-directory-original images> --img_gen_dir <path-to-image-directory-generated-images> --mode <mode>

Download pretrained checkpoint here.

@inproceedings{maharana2021integrating,
  title={Integrating Visuospatial, Linguistic and Commonsense Structure into Story Visualization},
  author={Maharana, Adyasha and Bansal, Mohit},
  booktitle={EMNLP},
  year={2021}
}