OpenImagingLab/AnyRecon

[SIGGRAPH Asia 2026] AnyRecon: Arbitrary-View 3D Reconstruction with Video Diffusion Model

What it solves

AnyRecon addresses the challenge of arbitrary-view 3D reconstruction. It allows users to generate high-quality 3D reconstructions from a set of capture views, enabling the creation of videos that show a scene from any desired camera trajectory.

How it works

The system uses a video diffusion model (specifically Wan2.1 I2V) combined with LoRA weights to refine and reconstruct 3D scenes. It relies on a point-cloud rendering pipeline (π³) to create condition videos from capture views. These condition videos, and the estimated or custom camera poses, guide the video diffusion model to generate the consistent 3D reconstruction across various views.

Who it’s for

This project is for researchers and developers working in 3D computer vision, neural rendering, and generative AI, specifically those looking to integrate video diffusion models for 3D scene reconstruction.

Highlights

  • Arbitrary-View Reconstruction: Generates 3D scenes that can be viewed from any angle.
  • Video Diffusion Integration: Leverages the pre-trained Wan2.1 I2V model for high-fidelity visual synthesis.
  • Custom Trajectories: Supports both pose estimation from test-view images and manually specified camera trajectories.
  • Point-Cloud Guided: Uses π³ for initial point-cloud rendering to provide structural guidance to the diffusion process.

Related

  • Project
    NVlabs/AnyFlow

    AnyFlow is an any-step video diffusion framework that allows a single model to adapt to arbitrary inference budgets for high-quality video generation.

  • Project
    WHU-USI3DV/VistaDream

    VistaDream is a training-free framework that reconstructs high-quality 3D scenes from a single-view image by ensuring consistency across generated novel views.

  • Project
    IamCreateAI/NeoVerse

    NeoVerse is a 4D world model that reconstructs 3D scenes from monocular videos to enable high-quality video generation along novel camera trajectories.

  • Project
    Eyeline-Labs/Vista4D

    Vista4D is a video reshooting framework that uses 4D point clouds to synthesize dynamic scenes from source videos using novel camera trajectories and viewpoints.

  • Project
    facebookresearch/mvdust3r

    A single-stage 3D scene reconstruction tool that creates 3D point clouds and camera poses from sparse RGB views in about 2 seconds without requiring pre-known camera poses.