How to solve the problem of face consistency when generating videos with Open-Sora?

2025-08-25

1.3 K

Specific ways to optimize facial consistency

Open-Sora, an open source video generation tool, provides the following solutions in terms of character face consistency:

Using the Image to Video Pipeline: Compared to direct text generation, outputting a high-quality character facial image at the text generation step before feeding it into the video generation process as a reference image can significantly improve facial stability.
Adjustment of sports scoring parametersThe -motion-score parameter is controlled to be between 3 and 5 (default 4) to avoid too high a value leading to facial deformation. Tests show that the most natural facial changes occur when motion-score=3.
reference frame control technology: Add the -keyframe-interval parameter (e.g., set to 10) to the i2v generation command to force the model to periodically reference the original facial features
High Resolution PriorityThe 768px model retains more facial detail than the 256px model 37%, and it is recommended to prioritize the use of the high resolution mode
Business Edition Program: The Video Ocean product offers enhanced face consistency algorithms and can be considered as a complementary solution.