Best Practices
How can I get the best results in DeepEditor?
Ensure your footage meets the following conditions:- Head rotation: We can see both eyes of the speaking character at all times.
- Head tilt: We can see the entire face. No extreme angles.
- Jaw in shot: We can see the entire face. No extreme closeups.
- Facial hair: Clean face to short facial hair. No full beards.
- Occlusions: No objects come in front of the lower half of the face.
- Lighting: All facial features are light enough to be distinguishable.
- Distortion: No unusual appearances (e.g. heavy make-up or non-human).
How can I reduce render times in the Refine tool?
To minimize render times in the Refine tool, ensure interpolation ramps are set correctly so that only the required frames get rendered. This means bringing interpolation down to 0% on frames that do not need to be rendered (e.g. where the source footage can be used) and using a transition of around 4 frames when increasing or decreasing interpolation values.
What are the refine tool best practices?
Stick to the Source Media as much as possible
If the vub is slightly out of sync, an NLE audio edit or nudge is a good starting point.
For example, if the character only speaks at the beginning of the shot, ease the Interpolation from 100 to 0 after the line.
Use a “light touch”
The goal is to minimize the addition of numerous keyframes to the timeline, as DeepEditor should automatically generate most of the final result.
Heavy movements can look unnatural. Start with 10 or -10. Blendshape adjustments should rarely go above 30 or -30.
Don’t overwork the shot. Too many refinements can make the output media look worse, not better.
Transition smoothly
One or two frames is not enough time for a mouth to completely change. “Ride the wave” and try to take at least 4 frames to transition.
Note: This is not a “rule”. Sometimes a character may speak very quickly and need sudden changes.
Light moves faster than sound
Starting a movement a frame or two early may look better to the viewer.