NeRF and 3DGS encode a scene with complementary inductive biases, yet existing cross‑representation distillation usually fixes one model as the teacher for the whole scene, which can spread local reconstruction errors. RouteBridge introduces a bidirectional framework that selects the teaching direction per ray. Its reliability estimator fuses photometric residuals with representation‑specific geometric evidence, routing supervision from NeRF to 3DGS, from 3DGS to NeRF, or abstaining when confidence is low. A renderer‑independent interface transfers color, opacity, and normalized depth without requiring shared features or point correspondences. On the mip‑NeRF 360 benchmark, the exported NeRF and 3DGS achieve PSNRs of 28.56 dB and 28.77 dB respectively. The 3DGS export improves over the original 3DGS by 1.56 dB and over NeRF‑GS by 0.45 dB while reducing LPIPS to 0.207. On the static three‑view DTU dataset, RouteBridge reaches 21.12 dB. Ablation studies confirm that both adaptive routing and geometric ray targets contribute to the gains.
Review