GODIVA: Generating Open-DomaIn Videos from nAtural Descriptions

TL;DR

GODIVA is a large-scale open-domain text-to-video model pretrained on extensive data, capable of generating videos from natural language with zero-shot abilities and evaluated using a new metric, addressing previous limitations in dataset size and generalization.

Contribution

This work introduces GODIVA, a novel open-domain text-to-video pretrained model utilizing 3D sparse attention and trained on the large-scale Howto100M dataset, enabling improved generalization and zero-shot performance.

Findings

GODIVA can generate videos from text with good zero-shot performance.

The model is pretrained on over 136 million text-video pairs.

A new metric, Relative Matching (RM), effectively evaluates video generation quality.

Abstract

Generating videos from text is a challenging task due to its high computational requirements for training and infinite possible answers for evaluation. Existing works typically experiment on simple or small datasets, where the generalization ability is quite limited. In this work, we propose GODIVA, an open-domain text-to-video pretrained model that can generate videos from text in an auto-regressive manner using a three-dimensional sparse attention mechanism. We pretrain our model on Howto100M, a large-scale text-video dataset that contains more than 136 million text-video pairs. Experiments show that GODIVA not only can be fine-tuned on downstream video generation tasks, but also has a good zero-shot capability on unseen texts. We also propose a new metric called Relative Matching (RM) to automatically evaluate the video generation quality. Several challenges are listed and discussed as future work.