# Meta and Oxford Researchers Introduce AVM, an Autoregressive Model for Free-Camera Scene Exploration

_Research · published 2026-11-20_

Meta and University of Oxford researchers have unveiled AVM (Autoregressive View Model), a generative model that represents 3D scenes as a stream of views—without building any explicit 3D representation such as a point cloud or mesh. Like a large language model, AVM prefills spatial prompts drawn from images, video, or multi-camera rigs, then autoregressively generates queried "view" tokens one block at a time, enabling free camera control and dynamic scene rendering across long rollouts.

The model accepts heterogeneous spatial prompts—single images, monocular video, or synchronized multi-camera rigs—and uses camera poses and timestamps to align observations into a shared scene context. Demonstrated capabilities include novel-view synthesis from sparse image sets, re-framing of live performances from virtual camera paths, and consistent geometry on out-and-back trajectories. The researchers highlight potential applications in virtual cinematography, sports broadcasting, and VFX production, as well as a design intended to scale toward real-time deployment.

## Sources
- [autoregressive-view-model.github.io](https://autoregressive-view-model.github.io/main.html)
---
Canonical: https://genbuzz.news/posts/meta-and-oxford-researchers-introduce-avm-an-autoregressive-model-for-free-camera-scene-exploration
