FlexAvatar turns a single portrait into a full 3D head in 2 minutes

Turning a single selfie into a full 3D avatar usually ends in heartbreak: the face looks decent straight-on, but turn the camera around and you are suddenly staring at a flat cardboard cutout with a missing skull.

FlexAvatar, a new research project headed to CVPR, wants to fix that messy tradeoff. The system takes one standard portrait image and reconstructs an animatable, complete 360-degree 3D head in just two minutes.

Untangling facial expressions from camera angles

Most single-camera avatar pipelines fall apart because they mix up two things: head rotation and facial movement. When someone smiles while turning, single-view models get confused about what changed in 3D space versus what just deformed on the surface.

FlexAvatar addresses this with a transformer-based encoder-decoder design paired with 3D Gaussian Splatting (3DGS) and DINOv2 vision representations. The encoder maps a single picture into a latent space, which decodes directly into 3D Gaussians for real-time rendering.

The real breakthrough lies in what the team calls "bias sinks"—learnable data tokens that bridge monocular videos and multi-view captures during training. This lets the network absorb broad face variations from internet clips while maintaining true volumetric depth from controlled multi-camera rigs.

Once ready, the model decouples viewpoint controls from facial expression inputs. You can spin around the head while making the avatar wink or speak, without tearing the mesh or smearing textures.

2 minutes to build, 1.7GB to run

The reconstruction speed is impressive, but the runtime footprint might be the bigger deal. Generating a fresh avatar takes only two minutes.

During inference, FlexAvatar delivers 20 FPS on an RTX 3090 while consuming a modest 1.7GB of VRAM. It also natively supports multi-shot inputs for better detail and allows smooth latent-space interpolations between different faces.

Why it matters

We think FlexAvatar stands out because it sidesteps the usual brute-force trap in 3D vision. Instead of demanding massive studio capture rigs for every new face, it relies on clever token design to reconcile messy web video with clean 3D supervision.

For game modders, indie creators, and XR developers, this radically cuts down the cost of populating worlds with believable digital humans. Anyone with a smartphone selfie can step into a scene without looking like a floating mask.

Its 1.7GB memory footprint is a massive practical win. It means real-time avatars can easily live alongside game engines on regular consumer GPUs, rather than monopolizing high-end cloud instances.

The caveat is that 20 FPS is still right on the edge of conversational latency, falling short of a buttery 60 FPS standard. Extreme lighting or occluded hair from single snapshots will also remain tricky edge cases until the codebase faces broader public testing.