✦
Summary
AnyRecon is a 3D reconstruction framework that builds high-quality scenes from arbitrary and unordered sparse views. It is designed for situations where users only have casual captures or limited input frames, but still need scalable reconstruction over large scenes and long camera trajectories. This makes it useful for 3D scanning, scene digitization, virtual production, and spatial AI research.
Key Features
Reconstructs 3D scenes from arbitrary sparse input views.
Supports unordered captures instead of requiring a rigid input sequence.
Uses video diffusion modeling for reconstruction and novel-view synthesis.
Maintains a persistent global scene memory for long-range conditioning.
Supports long trajectories with large numbers of generated frames.
Preserves explicit geometric control during reconstruction.
Useful for casual capture, 3D scanning, and scene digitization.
Provides public code and checkpoints for research evaluation.
About AnyRecon
The system combines explicit geometric control with video diffusion modeling to improve novel-view synthesis and reconstruction consistency. It uses a persistent global scene memory built from capture views, allowing the model to condition on more than one or two frames and maintain frame-level correspondence over major viewpoint changes. This helps address the scalability limits of earlier diffusion-based reconstruction methods.
AnyRecon is valuable because it makes sparse-view reconstruction more flexible and practical. It can support long trajectories with hundreds of frames, handle unordered inputs, and produce large-scale 3D reconstructions that are better suited to real capture conditions.