Meta can now turn 4 selfies into a photorealistic 3D avatar

If you wanted a photorealistic digital avatar in the past, you basically had two bad choices: sit inside a sci-fi rig packed with over 100 cameras, or hold your phone up, awkwardly make 65 facial expressions over three minutes, and let four top-tier GPUs cook the data for hours.

Now, Meta and the Technical University of Munich have published a paper introducing Avat3r, a large reconstruction model that cuts that entire ordeal down to just four smartphone selfies and a few minutes of processing.

From 100 cameras to four quick snaps

Meta has been pouring R&D into photorealistic avatars for over six years. Its top-shelf systems look incredible and cross the uncanny valley, but they require wildly expensive, studio-grade capture hardware that no normal consumer will ever touch.

Their previous attempts to simplify this down to smartphone cameras still felt like a chore. Users had to pose through 65 distinct expressions, and the processing took hours on beefy workstation hardware.

Avat3r changes the math by relying on a vision transformer (ViT). Just like large language models predict the next token in a sentence, Avat3r's transformer predicts a set of 3D Gaussians directly from those four 2D photos. If you've been following 3D Gaussian splatting in tech like Varjo Teleport or Meta's Horizon Hyperscapes, it is the same visual magic applied to human heads.

There is a massive catch, though: rendering performance. While building the model only takes a few minutes, rendering the animatable avatar currently chugs along at just 8 frames per second on an Nvidia RTX 3090. The animation system is also not yet hooked directly into VR headset sensors, though the authors note adapting face and eye tracking inputs should be straightforward.

Why it matters

For everyday users and VR enthusiasts, this is the first glimpse of a realistic setup pipeline that normal people might actually complete. Asking someone to snap four selfies is frictionless; asking them to sit through a complex calibration routine kills adoption. That said, with an 8 FPS output on high-end hardware, this is not something you will be booting up on your Quest anytime soon.

For the industry, Avat3r shows that generative 3D priors can replace exhaustive geometric scanning. Instead of capturing every wrinkle with dozens of physical cameras, a foundation model can "hallucinate" the missing angles accurately based on learned facial structures. This approach pushes digital human pipelines away from traditional photogrammetry rigs and toward large visual reconstruction models.

We think Meta's bet on shifting the burden from the user to the model is the right play. Generation time and capture friction are the biggest barriers to mainstream digital identities. An 8 FPS bottleneck on an RTX 3090 looks grim on paper, but inference optimization and Gaussian splatting renderers improve at breakneck speed. Once the rendering catches up, setting up a life-like avatar might finally take less time than creating a Mii.