Your head as a NeRF: drivable avatars from single videos

What it does

INSTA is a research implementation that turns video footage of a human head into a volumetric neural avatar. It sits on top of NVIDIA’s instant-ngp and incorporates a parametric FLAME mesh to anchor the NeRF representation, letting you render and reenact a specific person’s head from new viewpoints or expressions. The project ships with preprocessed datasets and checkpoints for several public avatar collections.

The interesting bit

The hardware split is striking: training demands an RTX 3090 and 32 GB RAM, but once exported to a snapshot the same avatar renders comfortably on an RTX 3080 laptop GPU. That separation of training cost from inference weight is unusual in the neural avatar space, where both phases typically demand top-tier hardware.

Key highlights

• Built directly on instant-ngp, inheriting its fast volumetric rendering
• Integrates FLAME mesh raycasting (toggle with F in the viewer) for structured head geometry
• Supports cross-reenactment via deformation gradient transfer
• Provides preprocessed avatars from NHA, NeRFace, and IMAvatar datasets
• Includes a separate PyTorch demo implementation for those who prefer Python over C++

Caveats

• Several features from the original instant-ngp codebase are unavailable here
• Training requires high-end hardware (RTX 3090, 32 GB RAM) and roughly 1,000 input frames
• Dataset preparation depends on external tracking repositories and a conda environment

Verdict

Researchers and graphics developers working on neural head avatars should look here, especially if they already live in the instant-ngp ecosystem. Casual developers or anyone hoping for a lightweight, self-contained face-swap tool should look elsewhere.