A method for accelerating a video diffusion model using synthetic datasets

By constructing a synthetic dataset and using knowledge distillation training with teacher and student models, the problems of time-consuming and unstable quality in the video diffusion model generation process were solved, achieving efficient, high-speed, and high-quality video generation.

CN120297362BActive Publication Date: 2026-06-05SHANGHAI ARTIFICIAL INTELLIGENCE INNOVATION CENT
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
SHANGHAI ARTIFICIAL INTELLIGENCE INNOVATION CENT
Filing Date
2025-03-25
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

Existing video diffusion models require multiple inference steps and a large amount of computational resources to generate videos, and distillation techniques are difficult to apply effectively in video diffusion models, resulting in a time-consuming generation process and unstable quality.

Method used

A synthetic dataset containing synthetic videos and denoised trajectories is generated using a pre-trained video diffusion model. The dataset is trained through knowledge distillation of the teacher and student models, and a trajectory-based loss function and adversarial training strategy are used to reduce inference steps and improve generation quality.

Benefits of technology

It significantly accelerates the video generation process, reduces inference steps, and improves the quality and resolution of generated videos, increasing generation speed by 8.5 times. The video quality performs excellently in the overall evaluation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120297362B_ABST
    Figure CN120297362B_ABST
Patent Text Reader

Abstract

The application discloses a method for accelerating a video diffusion model using a synthetic data set. The method comprises: generating a synthetic data set using a pre-trained video diffusion model, the synthetic data set containing a synthetic video, a denoising trajectory in a latent space, and a corresponding text prompt; using the pre-trained video diffusion model as a teacher model and constructing a corresponding student model, the student model and the teacher model sharing the same structure; based on the synthetic data set, performing knowledge distillation training on the student model, in the knowledge distillation training process, the student model learns the denoising process of the teacher model and aligns the data distribution generated by the teacher model until a set loss function standard is met; and using the knowledge distillation trained student model as a video generation model and applying it to a video analysis task. By using the application, a video with higher quality and higher resolution can be generated.
Need to check novelty before this filing date? Find Prior Art