LangFlash: Feed-forward 3D Language Gaussian Splatting from Sparse Unposed Images

Researchers introduced LangFlash, a feed-forward 3D reconstruction system that works from sparse, unposed multi-view images.

It predicts Gaussian-based geometry and aligned semantic features, enabling both novel view synthesis and language-aware scene understanding.

The method uses a semantically enriched RealEstate10k dataset plus a compact semantic encoding scheme to keep the model efficient.

Results show strong reconstruction quality and improved semantic consistency, pointing to faster pose-free 3D scene modeling for interactive multimodal systems.