ReconSplat achieves generalizable 3D reconstruction via diffusion priors

zhenjun_zhao · x · 2026-09-01

ReconSplat is a feed-forward model for 3D scene reconstruction that balances plausible view generation with geometric consistency. It integrates 3D Gaussian Splatting with a multi-view latent diffusion model, guiding the process with variational 3D latents. It achieves photorealistic novel views and accurate depth maps on real-world benchmarks.

Original post →

More from Research

Research channel →