LagerNVS: Latent Geometry for Fully Neural Real-time Novel View Synthesis
September 23, 2026
decoder-only architectures take a source view and target camera and output the target view.
-expensive because the entire network is re-evaluated for each new view
encoder-decoder architectures extract a scene's latent 3D representation once, while the decoder runs for every new view to synthesize.
-the encoder runs only once per scene
-bottleneck: latent tokens are constrained to certain dimensions before decoding
-highway: the decoder can directly access all image features rather than only one compressed 3D latent