A photograph contains one view of a person at one moment. An animated head needs many views and many moments. The gap between those two is where a neural avatar model does most of its work.
MegaPortraits explores high-resolution head avatars, including the difficult case where one person's motion drives a different source identity. The July paper combines information from medium-resolution video and high-resolution images. It also describes a distilled real-time model whose available identities are restricted to a predefined set.
Those are distinct capabilities. A model that can generate a new identity from one source image and a faster specialized model with fixed identities do not have identical deployment properties. The distinction matters when imagining an interactive product.
What can move across people?
A driving video contains head motion, expression, lighting and the driver's own facial proportions. The desired transfer usually includes some of those signals and excludes others. If the driver opens their mouth, the source should speak or react. The source should not quietly acquire the driver's face shape.
That separation is difficult because the evidence is mixed together in the image. A dark region near the cheek could be a shadow, a feature of the face or a consequence of rotation. The model must infer a useful explanation from patterns learned across other people and motions.
The source photograph also leaves much unseen. A frontal portrait provides limited evidence about the ears, the side of the head and what becomes visible during a large turn. A convincing output can therefore contain plausible invention. It should not be treated as a recovered recording of how that person actually looks from every angle.
Resolution makes both the success and the failure more visible. More pixels can preserve fine appearance cues, but they also make inconsistencies around teeth, eyes and hair easier to inspect. A high-resolution image that changes identity slightly from frame to frame may feel less convincing than a simpler but stable one.
For an avatar, I would evaluate identity and motion separately. Does the result still resemble the source during a turn? Does the intended expression remain readable? Does the driver introduce unwanted appearance? A single attractive clip cannot answer every part of that comparison.
The transition back to neutral is especially informative. Large expressions are expected to change the face. Small residual movements after the expression ends reveal whether the representation can settle coherently. Hair and background boundaries can also expose changes that are easy to overlook when watching the mouth.
I can imagine useful applications in stylized communication, controlled performances and character experiments. Their requirements differ. A playful character can tolerate interpretation that would be unacceptable in a representation intended to document a real person's appearance.
What interests me is the deliberate division between identity and motion. It is a powerful interface: one image supplies who, another stream supplies how they move. The quality of the result depends on how carefully the system keeps those responsibilities separate, and how clearly the application presents the parts it has had to invent.