Skip to content

ReconFusion

Rebuild real scenes from a handful of photos with diffusion priors

Visit official site

What ReconFusion is

The method is described as a 3D reconstruction approach that uses diffusion priors with Neural Radiance Fields to rebuild real world scenes from a small set of photos.

What you can do with it

  • Rebuild a scene from a few views
  • Extend geometry to new camera positions
  • Keep the appearance of observed regions

How it works

  • NeRF optimisation reduces reconstruction and sample loss
  • A PixelNeRF style model generates sample images
  • Noisy latents feed a diffusion model

Pros & cons

✓ What we like

  • Geometry and texture extend beyond input camera positions
  • Observed regions keep their appearance

! What to watch out for

  • Training relies on few-view and multi-view datasets
  • The approach is built around NeRF optimisation

FAQ

How few images are needed?

The material describes reconstruction from only a small number of photographs.

What is combined with NeRF?

The material names diffusion priors and a PixelNeRF style model.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review