Training Data and Model Training
Learn about model training and training data.
What is Training Data?
How can I get the best results when using Training Data?
How much Training Data can I upload?
Does Flawless train AI models using my uploaded footage?
What is Training Data?
Training Data is additional footage of the character whose dialogue you are altering, uploaded as reference material for DeepEditor to analyze and produce a more precise Vub. The more training data, the better the result, provided the additional footage complies with the guidelines below.
How can I get the best results when using Training Data?
To get the best results use footage that is really similar to your Source media (e.g. alternate takes) and meets the following criteria:
- The person is speaking and their mouth interior is visible (this helps DeepEditor learn the way the person's mouth moves when they speak)
- The person's face position and angle is very similar to the Source media
- Lighting and background are very similar to the Source media
- There are at least 15 seconds of footage showing a variety of mouth shapes (footage can be in separate video clips)
- Each clip features the same actor as the Source media
- The person's face is clear and unobstructed
- The media is compliant with the DeepEditor Supported Media Formats
How much Training Data can I upload?
Currently, you can upload up to 40 video files of training data of which you can select up to 5000 frames to be used towards training. Individual files should be shots not containing cuts. The frame limit for an individual video file is 1440.
Note that Vubs with Training Data will take significantly more time to render and could take over 24 hours to complete, depending on the amount of Training Data you upload.
Does Flawless train AI models using my uploaded footage?
No, Flawless does not use customer data to train models. See more on copyright and legal agreements concerning uploaded footage here.