GeoVerse is a framework for novel view synthesis from sparse images. It addresses the tradeoff between faithfully reconstructing observed regions and plausibly completing unseen content while maintaining world consistency across viewpoints. The method works by embedding appearance priors from a video (Wan2.2 VACE) into the geometric latent space of a pretrained 3D foundation model, using a ControlNet-style adapter. A global spatial memory enforces cross-view coherence by continuously aggregating observed and synthesized content and reprojecting it as pixel-aligned guidance for subsequent views.
On benchmarks, GeoVerse achieves a 2.23 dB higher PSNR on DL3DV and a 32.4% lower ATE (pose error) on Mip-NeRF360 compared to the previous state-of-the-art Geometry Latent Diffusion (GLD). The framework also scales training from 4 to 15 datasets and uses reflow distillation to reduce denoising steps to 4, resulting in inference under 10 seconds—a 17× speedup over GLD.
This work is interesting because it provides a practical way to combine the geometric consistency of 3D reconstruction models with the generative richness of video diffusion models, achieving both high visual quality and spatial coherence without the drift typically seen in sequential video-based approaches. It offers a strong foundation for building multi-view consistent world models.