ABSTRACT

Neural Radiance Fields (NeRF) have revolutionized photorealistic 3D synthesis but are often limited by prohibitive storage requirements. We propose HyBiC, a novel compression framework that integrates a Hybrid Binary-Real representation with advanced contextual entropy modeling. While building on the efficiency of binary grids, our method overcomes the inherent precision limits of full binarization by introducing a real-valued 3D coarse grid. This hybrid hierarchy preserves critical geometric structures and enhances scene expressiveness without the storage overhead of traditional volumetric grids. To maximize coding efficiency, we introduce a grouped channel-wise autoregressive context model that explicitly targets both spatial and inter-channel redundancies. Experimental results demonstrate that HyBiC achieves a new state-of-the-art (SOTA) in NeRF compression, offering a balance between fidelity and storage efficiency. 

PROPOSED METHOD

We propose HyBiC, a compression framework that achieves high-fidelity scene reconstruction with minimal storage. The overall architecture of HyBiC is illustrated in Fig. 1. Our method is built upon two core pillars: a hybrid feature representation that synergies structural precision with memory efficiency, and an advanced context model designed to maximize bit-rate reduction by exploiting feature dependencies. 

A. Hybrid Feature Representation

We decompose the scene representation into a dual-level hierarchy: a coarse level modeled by 3D grids and a fine level represented by 2D grids. This differs from BiRF, which employs a more redundant structure by interleaving both 3D and 2D grids across multiple layers. To achieve compact storage and high expressive capacity, we construct our feature grids based on the multi-resolution hash encoding introduced in in Instant-NGP. Hash encoding mechanism allows us to define extremely high-resolution grids to capture fine details without being constrained by available memory. 

To efficiently capture high-frequency details, we employ multi-resolution 2D feature grids with binary parameters for the fine level representation. This strategy significantly reduces the memory compared to 3D volumetric storage. We adopt a binary encoding scheme where real-valued parameters θ are quantized to θ ′ ∈ {+1, −1}: θ ′ = sign(θ) = ( +1 if θ ≥ 0 −1 if θ < 0. As the binarization process is inherently non-differentiable, we utilize the Straight-Through Estimator (STE) to approximate gradients during backpropagation. 

At the coarse levels, we employ multi-resolution 3D grids to capture the global scene structure. We retain these as realvalued parameters, as binary values lack the expressive capacity required for high-fidelity global geometry. To mitigate the resulting storage overhead, our context model (Sec. 3.3) leverages hierarchical dependencies to efficiently compress these layers. This approach strikes an optimal balance between structural precision and bit-rate efficiency. 

B. Feature Query and Rendering

    In the fine-level representation, we employ a triplane decomposition consisting of three 2D multi-resolution feature grids. Given a 3D coordinate, the associated 2D features are retrieved by projecting x onto each 2D planes and performing bilinear interpolation. For example, the feature queried from the yz-plane. 

Similarly, we can query the features from fxy and fxz. We aggregate the features from the multi-resolution tri-plane grids to form the final 2D feature representation Similarly, for the coarse-level 3D grids, we utilize trilinear interpolation to extract multi-resolution features. Specifically, given a spatial query x = (x, y, z), we sample the corresponding feature vectors from each resolution level l (ranging from 1 to L) and aggregate them to construct the 3D feature representation. 

The retrieved features from both 3D and 2D grids are concatenated into a single vector and fed into a density MLP to predict spatial density and intermediate features. The intermediate features are then passed to a color MLP to predict the RGB color of x. 

C. Compression

    Our entropy coding module then builds upon the CNC framework [11], which effectively eliminates spatial and hierarchical redundancies. To enable entropy coding, the continuous feature values in our 3D coarse grid must first be discretized. We employ uniform scalar quantization and the quantized features. To allow for end-to-end optimization during training, we simulate the quantization error by adding uniform noise to the feature values. 

Channel-wise Autoregressive model. For a quantized feature vector at location, CNC derives a spatial context vector vsp from the coarser-level parents. While this minimizes inter-level dependencies, CNC treats the C channels independently. We argue that this process is suboptimal for real-valued NeRF features, which exhibit strong inter-channel correlations. Neglecting these dependencies leaves residual redundancy that limits the compression ceiling. 

To address this, we introduce a grouped channel-wise auto-regressive mechanism that balances modeling capacity and inference throughput. We factorize the joint distribution of the feature vector, conditioned on spatial context and previously decoded channels, into a product of conditional probabilities.

For the 2D fine grids, we adopt a 3D-to-2D context modeling approach from CNC. Recognizing that the highestresolution 3D grid contains the most comprehensive spatial priors, we project its voxel features along the three primary axes. These projections serve as auxiliary spatial contexts for the 2D triplanes, guiding the probability estimation and further enhancing the compression of the fine-level representation. 

EXPERIMENT

A. Dataset

We evaluate our method on the NeRF-Synthetic dataset. To accelerate ray marching and rendering, we leverage the occupancy grid estimator provided by NerfAcc 

B. Effectiveness of Real-Valued Hybrid Representation

    To validate the advantage of our real-valued strategy, we compare HyBiC-R against the fully binary baseline BiRF, across varying feature dimensions F = {1, 2, 4, 8}, as shown in Table 1. We denote these variants as BiRF-F and Ours-F while Ours-8* is used in HyBiC-C. Table 1 shows that our approach consistently outperforms BiRF [7] in reconstruction quality while maintaining a smaller storage. For F = 8, our model achieves a PSNR of 33.71 dB, surpassing BiRF-8 (33.59 dB), while simultaneously reducing the model size from 5.8 MB to 4.4 MB. This confirms that incorporating real-valued features allows for a more expressive representation, enabling us to achieve higher fidelity with a more compact memory footprint. 

C. Rate-Distortion Performance

    We evaluate the R-D performance of HyBiC-C against state-of-the-art compression frameworks, as shown in Fig. 2, For this comparison, Ours-8*, representing a configuration with 11 levels of 3D+2D grids and 8 features perlevel is served as the input for HyBiC-C. Fig. 2 demonstrates that our method consistently outperforms all existing baselines, delivering superior reconstruction quality across all bitrates. Notably, even at lowest rate, HyBiC-C achieves a high PSNR of 33.75 dB with only 0.78 MB of storage. This highly compressed result is still higher than the peak reconstruction quality of several baselines that require substantially more storage. Furthermore, we compare the variants BiRF-F, Ours-F, and the R-D optimized Ours-8. The superior performance of HyBiC-C highlights that achieving an optimal R-D trade-off depends not only on entropy coding but also on the inherent expressiveness of the underlying hybrid representation. To investigate the source of our compression gains, we analyze the storage distribution of Ours-8* (the configuration used in HyBiC-C) before and after compression. As illustrated in Figure 3, the total storage requirement reduces from 22.05 MB to 1.11 MB, achieving an impressive 20X compression ratio. The most significant reduction occurs in the coarse level, which shrinks from 21.161 MB to 0.504 MB. While uncompressed coarse grids typically dominate the storage budget due to dense floating-point parameters, our entropy coding framework effectively eliminates this redundancy. In the final