RT-NeRV: Rethinking Hybrid Neural Representations for Video via Residual Tokenization

Neural Representations for Videos(NeRV) have emerged as a promising paradigm for video compression by representing videos as compact neural networks with efficient decoding. Hybrid NeRV methods further improve reconstruction quality through content adaptive embeddings, but still struggle to preserve fine details at low bitrates. A key limitation is that shallow residual support in formation, although highly beneficial for reconstruction, is costly to transmit in its continuous form and is therefore underutilized. In this paper, we rethink hybrid NeRV and present RT-NeRV, a residual tokenization framework for hybrid neural video representations. The core idea is to discretize shallow residual features and inter-frame residual cues into compact residual tokens, allowing informative reconstruction support to be transmitted efficiently and exploited by the decoder. To this end, we design a residual tokenizer together with a residual-aware codebook learning strategy that improves token utilization and stabilizes training. RT-NeRV can be readily integrated into modern hybrid NeRV hosts, consistently enhancing detail preservation, reconstruction quality, and bitrate quality trade-offs. Extensive experiments on video regression and related restoration tasks show that RT-NeRV outperforms strong hybrid NeRV baselines and remains competitive with recent INR based video compression methods. These results demonstrate that residual tokenization is an effective and complementary direction for advancing hybrid neural video representations