MAGiC-NeRF: Multimodal Generative Priors and Conditional Flow for Extreme NeRF Compression

MaGic-NeRF introduces an extreme NeRF compression framework that distills trained radiance fields into compact triplanes and reconstructs them from highly compressed multimodal semantic priors. By combining flow-based triplane recovery with agent-guided image decomposition, it achieves very high compression ratios while preserving rendering quality across indoor, object-level, and real-world scene benchmarks.

Recommended citation: Xiaoge Zhang, Zijie Wu, Mingtao Feng, Lei Yang, Zichen Geng, Saeed Anwar, & Ajmal Mian. (2026). MAGiC-NeRF: Multimodal Generative Priors and Conditional Flow for Extreme NeRF Compression.