Invention Title:

GENERATING MULTI-SUBJECT AND CUSTOMIZED MOTION VIDEOS

Publication number:

US20260195932

Publication date:
Section:

Physics

Class:

G06T11/00

Inventors:

Applicant:

Smart overview of the Invention

The patent application introduces a novel system for generating customized text-to-video outputs, focusing on high-quality videos that incorporate specified identities and motion patterns. It utilizes an appearance-agnostic motion learning approach to separate motion patterns from appearance features. A spatial-temporal collaborative composition scheme is employed to integrate learned multi-subject and motion LoRAs, enhancing the flexibility and generalizability of text-to-video generation.

Technical Details

The proposed system includes several components: a subject learner, a motion learner, and a spatial-temporal collaborative composer. The subject learner generates token low-rank adaptations (LoRAs) for each subject, while the motion learner creates a motion LoRA from reference motion videos. The spatial-temporal collaborative composer then combines these LoRAs to produce a coherent video output, integrating multiple subjects and motion patterns using spatial-temporal sampling.

Methodology

The method involves receiving images of subjects, a motion video, and a text prompt. A reference motion is learned through an appearance-agnostic motion learning process, guided by the text prompt. A spatial-temporal algorithm is applied to compose an initial video, which is then used to generate the final output video featuring the specified subjects and motion patterns.

System Architecture

The system architecture comprises a receiver for input parameters, a LoRA generator, and processors executing code to create the output video. The LoRA generator produces subject and motion LoRAs using the input data, while the processors utilize a spatial-temporal collaborative composition algorithm within a diffusion model framework to synthesize the video, ensuring precise control over noise and motion integration.

Innovative Framework

The disclosed processes provide a unified framework for video content customization, enabling control over subject identities and motion patterns. By employing subject and motion LoRAs, and utilizing appearance-agnostic motion learning alongside spatial-temporal composition, the system enhances user control and coherence in video generation, supporting complex multi-subject interactions and dynamic motions.