Understanding the value of variety

Variety is the single most defining trait of a good dataset. The best datasets will all have a large variety of poses, expressions, and lighting situations, while the worst results will come from data lacking variety in one or more of these categories. Some of the areas of variety we’ll cover in this section include pose, expression, and lighting.

Pose

Pose is a simple category to both see and understand. Pose is simply the direction and placement of the face in the image. When it comes to deepfakes, only the pose of the face itself matters – the rest of the body’s pose is ignored. Pose in deepfakes is important so that the AI can learn all the angles and directions of the face. Without sufficient pose data, the AI will struggle to match the direction of the face and you’ll end up with poor results.

Figure 3.2 – Examples of different poses

One quick way to get the full range of...