The Hollywood at Home: DragNUWA is an AI Model That Can Achieve Controllable Video Generation

Generative AI has made a huge leap in the last two years thanks to the successful release of large-scale diffusion models. These models are a type of generative model that can be used to generate realistic images, text, and other data. 

Diffusion models work by starting with a random noise image or text and then gradually adding detail to it over time. This process is called diffusion, and it is similar to how a real-world object gradually becomes more and more detailed as it is formed. They are typically trained on a large dataset of real images or text. 

On the other hand, video generation has also witnessed remarkable advancements in recent years. It encompasses the exciting capability of generating lifelike and dynamic video content entirely. This technology leverages deep learning and generative models to generate videos that range from surreal dreamscapes to realistic simulations of our world.

The ability to use the power of deep learning to generate videos with precise control over their content, spatial arrangement, and temporal evolution holds great promise for a wide range of applications, from entertainment to education and beyond.

Historically, research in this domain primarily centered around visual cues, relying heavily on initial frame images to steer the subsequent video generation. However, this approach had its limitations, particularly in predicting the complex temporal dynamics of videos, including camera movements and intricate object trajectories. To overcome these challenges, recent research has shifted towards incorporating textual descriptions and trajectory data as additional control mechanisms. While these approaches represented significant strides, they have their own constraints.

Let us meet DragNUWA which tackles these limitations.

DragNUWA is a trajectory-aware video generation model with fine-grained control. It seamlessly integrates text, image, and trajectory information to provide strong and user-friendly controllability.

Example videos generated with DragNUWA. Source: https://arxiv.org/pdf/2308.08089.pdf

DragNUWA has a simple formula to generate realistic-looking videos. The three pillars of this formula are semantic, spatial, and temporal control. These controls are done with textual descriptions, images, and trajectories, respectively.

The textual control is done in the form of textual descriptions. This injects meaning and semantics into video generation. It enables the model to understand and express the intent behind a video. For instance, it can be the difference between depicting a real-world fish swimming and a painting of a fish.

For the visual control, images are used. Images provide spatial context and detail, helping to accurately represent objects and scenes in the video. They serve as a crucial complement to textual descriptions, adding depth and clarity to the generated content.

These are all familiar things to us, and the real difference DragNUWA makes can be seen in the last component: the trajectory control. DragNUWA uses open-domain trajectory control. While previous models struggled with trajectory complexity, DragNUWA employs a Trajectory Sampler (TS), Multiscale Fusion (MF), and Adaptive Training (AT) to tackle this challenge head-on. This innovation allows for the generation of videos with intricate, open-domain trajectories, realistic camera movements, and complex object interactions.

Overview of DragNUWA. Source: https://arxiv.org/pdf/2308.08089.pdf

DragNUWA offers an end-to-end solution that unifies three essential control mechanisms—text, image, and trajectory. This integration empowers users with precise and intuitive control over video content. It reimagines trajectory control in video generation. Its TS, MF, and AT strategies enable open-domain control of arbitrary trajectories, making it suitable for complex and diverse video scenarios.


Check out the Paper and ProjectAll Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 30k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..

Ekrem Çetinkaya received his B.Sc. in 2018, and M.Sc. in 2019 from Ozyegin University, Istanbul, Türkiye. He wrote his M.Sc. thesis about image denoising using deep convolutional networks. He received his Ph.D. degree in 2023 from the University of Klagenfurt, Austria, with his dissertation titled "Video Coding Enhancements for HTTP Adaptive Streaming Using Machine Learning." His research interests include deep learning, computer vision, video encoding, and multimedia networking.

🐝 Join the Fastest Growing AI Research Newsletter Read by Researchers from Google + NVIDIA + Meta + Stanford + MIT + Microsoft and many others...