September 19, 2023

Google Introduces Innovative Generative Image Dynamics That Simulate Dynamic Scenes in Static Images

by Damir Yalalov

Published: September 19, 2023 at 4:21 am Updated: September 19, 2023 at 9:02 am

by Danil Myakin

Edited and fact-checked:

Google has unveiled a Generative Image Dynamics, a novel approach enables the transformation of a single static image into a seamless looping video or an interactive dynamic scene, offering a wide array of practical applications.

At the core of this pioneering technology is the modeling of an image-space prior on scene dynamics. The objective is to create a comprehensive understanding of how objects and elements within an image may behave when subjected to various dynamic interactions. This understanding can then be used to simulate the response of object dynamics to user interactions effectively.

The key feature of this technology is the ability to generate seamless looping videos. By leveraging the image-space prior on scene dynamics, Google’s system can extrapolate and extend the motion of elements within an image, transforming it into a captivating and continuous video loop. This functionality opens up numerous creative possibilities for content creators and designers.

The paper presents an approach to modeling an image-space prior based on scene dynamics, which is learned from a collection of motion trajectories extracted from real video sequences containing natural, oscillating motion such as trees, flowers, candles, and clothes blowing in the wind. The trained model uses a frequency-coordinated diffusion sampling process to predict a per-pixel long-term motion representation in the Fourier domain, which they call a neural stochastic motion texture. This representation can be converted into dense motion trajectories that span an entire video.

The technology enables users to interact with objects within static images realistically. By simulating the response of object dynamics to user excitation, Google’s system allows for immersive and interactive experiences within images. This has the potential to revolutionize metaverse spaces and how users engage with visual content.

The study explores modeling a generative prior for image-space scene motion, i.e., the motion of all pixels in a single image. The model is trained on automatically extracted motion trajectories from a large collection of real video sequences. Conditioned on an input image, the trained model predicts a neural stochastic motion texture: a set of coefficients of a motion basis that characterize each pixel’s trajectory into the future.

The foundation of this innovation lies in a meticulously trained model. Google’s model learns from a vast dataset of motion trajectories extracted from real video sequences featuring natural, oscillating motion. These sequences include scenes with elements like trees swaying, flowers moving, candles flickering, and clothes billowing in the wind. This diverse dataset enables the model to understand a broad range of dynamic behaviors.

The scope of the study is limited to real-world scenes with natural, oscillating dynamics, such as trees and flowers moving in the wind. The Fourier series is chosen as the basis functions. The resulting frequency-space textures can then be transformed into dense, long-range pixel motion trajectories, which can be used to synthesize future frames, turning still images into realistic animations.

When presented with a single image, the trained model employs a frequency-coordinated diffusion sampling process. This process predicts a per-pixel long-term motion representation in the Fourier domain, termed a neural stochastic motion texture. This representation is then transformed into dense motion trajectories that span an entire video. Coupled with an image-based rendering module, these trajectories can be harnessed for various practical applications.

Compared with priors over raw RGB pixels, priors over motion capture more fundamental, lower-dimensional under-dimensional structure that efficiently explains variations in pixel values. This leads to more coherent long-term generation and more fine-grained control over animations compared to prior methods that perform image animation via raw video synthesis.

The generated motion representation is convenient for a number of downstream applications, such as creating seamless looping videos, editing the generated motion, and enabling interactive dynamic images, simulating the response of object dynamics to user-applied forces.

Read more related topics:

Tags:

Disclaimer

Any data, text, or other content on this page is provided as general market information and not as investment advice. Past performance is not necessarily an indicator of future results.

The Trust Project is a worldwide group of news organizations working to establish transparency standards.

Damir is the team leader, product manager, and editor at Metaverse Post, covering topics such as AI/ML, AGI, LLMs, Metaverse, and Web3-related fields. His articles attract a massive audience of over a million users every month. He appears to be an expert with 10 years of experience in SEO and digital marketing. Damir has been mentioned in Mashable, Wired, Cointelegraph, The New Yorker, Inside.com, Entrepreneur, BeInCrypto, and other publications. He travels between the UAE, Turkey, Russia, and the CIS as a digital nomad. Damir earned a bachelor's degree in physics, which he believes has given him the critical thinking skills needed to be successful in the ever-changing landscape of the internet.

Damir Yalalov