GeoVerse: World-Consistent Novel View Synthesis in Geometric Latent Space
GeoVerse synthesizes world-consistent novel views inside a geometric latent space using video appearance priors.
GeoVerse synthesizes world-consistent novel views by generating inside the geometric latent space of a pretrained 3D foundation model while injecting appearance priors from a video model. Multilevel features from Wan2.2 VACE are added through a ControlNet-style adapter, and a global spatial memory reprojects observed and synthesized content to anchor later views. Compared with GLD, it reports 2.23 dB higher PSNR on DL3DV and 32.4% lower absolute trajectory error on Mip-NeRF360.
- Generates novel views inside a pretrained 3D model's geometric latent space.
- Injects Wan2.2 VACE appearance features through a ControlNet-style adapter.
- A global spatial memory reprojects guidance to keep viewpoints consistent.
- Gains 2.23 dB PSNR on DL3DV and cuts Mip-NeRF360 ATE 32.4% versus GLD.
From the original article
1 min read at huggingface.co
Novel view synthesis from sparse images must reconcile faithful reconstruction of observed regions with plausible completion of unseen content, while maintaining world consistency across viewpoints.
By Kerui Ren, Tao Lu, Linning Xu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.