Why Deepfakes Jitter on Turns, and How to Cure It
Jitter is not one problem but three, each cured differently. The first job is to work out which one you are looking at.
Type 1. Frame-to-frame flicker
The face as a whole changes shape or brightness slightly between frames. The cause is that each frame was processed independently and the model reached a slightly different decision each time.
The cure is temporal smoothing: the result is processed with reference to neighbouring frames rather than in isolation. The side effect is a slight loss of sharpness, so it is important not to overdo it.
Type 2. Edge jumps
The face itself is stable but the seam with its surroundings twitches. This is usually a detection error: on some frames the landmark points landed differently and the mask shifted.
The cure is smoothing the landmark trajectory over time and widening the mask. If the failures fall on specific frames, cutting that passage is easier than fighting it with settings.
Type 3. Drifting tone
The face gradually shifts to a different cast over the course of the clip. The cause is automatic colour correction adapting to each frame separately, or a change of lighting in the scene.
The cure is fixing colour parameters from a reference frame for the whole passage.
What matters most
The stability of the target footage. Shooting from a tripod gives a visibly cleaner result than handheld, even with identical processing settings. Camera shake compounds generation jitter, and visibility grows faster than the sum of the two.
Go to the site →Read next
Model Consent: Why There Is No Work Without It Neural Networks for Adult Content: What They Can Do Today How Image Generation Works: Diffusion in Plain Terms Adult Content Generation: What Models Can Do and Where the Limit IsThis material is intended for adults over 18 and is informational in nature.