Skip to content

[CVPR 2022] Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioning

License

Notifications You must be signed in to change notification settings

snap-research/MMVID

Repository files navigation

MMVID
Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioning (CVPR 2022)

Generated Videos on Multimodal VoxCeleb

This repo contains the code for training and testing, models, and data for MMVID.

Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioning
Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, Sergey Tulyakov
Snap Inc., Rutgers University
CVPR 2022

MMVID Code

CLIP model

Download OpenAI's pretrained CLIP model and place it under ./ (or any other directory that is consistent with arg --openai_clip_model_path),

wget https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt

VQGAN

Code for finetuning VQGAN models is provided in this repo.

Multimodal VoxCeleb

For testing, please download pretrained models and change the path for --dalle_path in the scripts.

For quantitative evaluation, append --eval_mode eval to each testing command. Output log directory can be changed by appending --name_suffix _fvd to add suffix (example here).

Text-to-Video

Training:

bash scripts/mmvoxceleb/text_to_video/train.sh

Testing:

bash scripts/mmvoxceleb/text_to_video/test.sh

For Quantitative Evaluation (FVD and PRD):

bash scripts/mmvoxceleb/text_to_video/evaluation.sh
Text Augmentation

Text augmentation for better training. To enable using a pretrained RoBERTa model, append --fixed_language_model roberta-large to the training/testing command. Note that this feature is only experimental and is not very robust.

To enable text dropout, append --drop_sentence to the training command. Text dropout is also compatible with using a RoBERTa. We observed that text dropout genrally improves diversity in the generated videos.

Training:

bash scripts/mmvoxceleb/text_augement/train.sh

Testing:

bash scripts/mmvoxceleb/text_augement/test.sh
Text and Mask

Training:

bash scripts/mmvoxceleb/text_and_mask/train.sh

Testing:

bash scripts/mmvoxceleb/text_and_mask/test.sh
Text and Drawing

Training:

bash scripts/mmvoxceleb/text_and_drawing/train.sh

Testing:

bash scripts/mmvoxceleb/text_and_drawing/test.sh
Drawing and Mask

Training:

bash scripts/mmvoxceleb/drawing_and_mask/train.sh

Testing:

bash scripts/mmvoxceleb/drawing_and_mask/test.sh
Image and Mask

Training:

bash scripts/mmvoxceleb/image_and_mask/train.sh

Testing:

bash scripts/mmvoxceleb/image_and_mask/test.sh
Text and Partial Image

Training:

bash scripts/mmvoxceleb/image_and_mask/train.sh

Testing:

bash scripts/mmvoxceleb/image_and_mask/test.sh
Image and Video

Training:

bash scripts/mmvoxceleb/image_and_mask/train.sh

Testing:

bash scripts/mmvoxceleb/image_and_mask/test.sh

Pretrained Models

Pretrained models are provided here.

Multimodal VoxCeleb

Weight FVD
VQGAN (vae) ckpt -
VQGAN (cvae for image conditiong) ckpt -
Text-to-Video pt 59.46
Text-to-Video (ARTV) pt 70.95
Text and Mask pt -
Text and Drawing pt -
Drawing and Mask pt -
Image and Mask pt -
Text and Partial Image pt -
Image and Video pt -
Text-Augmentation pt -

Multimodal VoxCeleb Dataset

Multimodal VoxCeleb Dataset has a total of 19,522 videos with 3,437 various interview situations (453 people). Please see details about how to prepare the dataset in mm_vox_celeb/README.md. Preprocessed data is also available here.

Acknowledgement

This code is heavily based on DALLE-PyTorch and uses CLIP, Taming Transformer, Precision Recall Distribution, Frechet Video Distance, Facenet-PyTorch, Face Parsing, and Unpaired Portrait Drawing.

The authors thank everyone who makes their code and models available.

Citation

If our code, data, or models help your work, please cite our paper:

@inproceedings{han2022show,
title={Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioning},
author={Han, Ligong and Ren, Jian and Lee, Hsin-Ying and Barbieri, Francesco and Olszewski, Kyle and Minaee, Shervin and Metaxas, Dimitris and Tulyakov, Sergey},
booktitle={CVPR},
year={2022}
}