Building lifelike 3D avatars usually means either rendering with explicit 3D geometry or collecting costly 3D capture data. The authors of this paper argue that video diffusion models, trained on internet-scale video, already encode a strong implicit understanding of appearance, motion, and physics. That prior is hard to replicate with traditional graphics pipelines.

The paper, titled Video Diffusion Avatars, introduces a way to tame such video diffusion models into controllable 3D avatars. Rather than learning everything from scratch, the approach taps into the knowledge already stored in the pretrained model. The abstract notes that this is an alternative to explicit 3D rendering and expensive data capture, though the full technical details are not available in the abstract alone.

This fits a broader pattern of adapting generative video models for 3D applications. If the method works as described, it could offer a cheaper path to personalized or animated avatars by leveraging priors learned from massive amounts of everyday video.