NVRC-Lite: Lightweight INR Video Compression

• NVRC-Lite is an ultra-lightweight variant of NVRC that employs an INR-based approach with a modified multi-scale HiNeRV design for efficient video compression.
• It leverages high-resolution multi-scale feature grids to compensate for reduced network complexity, achieving around 7–8 kMACs/pixel while maintaining high fidelity.
• Its octree-based entropy model reduces sequential decoding steps, resulting in significant speed improvements, including up to 8.4× faster encoding and 2.5× faster decoding compared to similar codecs.

NVRC-Lite is an ultra-lightweight variant of Neural Video Representation Compression (NVRC) for implicit neural representation (INR)-based video compression. It retains the core NVRC premise that a video can be represented by an over-fitted, video-specific neural function whose parameters are quantized and entropy-coded, but it reworks both the representation and the entropy model for the low-complexity regime. The defining modifications are a lightweight multi-scale HiNeRV-based representation with higher-resolution feature grids and an octree-based context model for entropy coding high-dimensional feature grids. In the reported experiments, these changes place NVRC-Lite at about $7$–$8$ kMACs/pixel while improving coding efficiency and speed relative to prior lightweight INR baselines such as C3 (Kwan et al., 3 Dec 2025).

1. Origins within INR-based video compression

NVRC-Lite emerges from the INR line of video codecs in which a neural representation $F_\theta$ is optimized directly for a single video sequence and the resulting parameters are compressed. In the NVRC framework, the representation parameters $\theta$, the quantization and entropy-model parameters $\phi$, and higher-level parameters $\psi$ used to code $\phi$ are organized hierarchically, enabling end-to-end rate–distortion optimization of both the representation and the compression machinery itself (Kwan et al., 2024).

NVRC-Lite preserves that overall philosophy but targets a different operating point. The paper explicitly positions it in the ultra-lightweight regime, where the central difficulty is not only rate–distortion performance but also computational practicality. Prior high-performance INR codecs can achieve strong compression but often do so with far higher decoding complexity, while prior lightweight INRs can be fast but give up compression performance. NVRC-Lite addresses that tension by redesigning NVRC around a small representation and a faster entropy coder rather than merely pruning a larger codec (Kwan et al., 3 Dec 2025).

A later development, NVRC++, makes this positioning even clearer by describing the broader design space as a scalability–performance trade-off: high-performance INR codecs such as NVRC can require increasing model complexity at higher quality, whereas lightweight codecs of the NVRC-Lite type can lose compression performance when scaled across bitrate and quality levels (Kwan et al., 26 Jun 2026). This suggests that NVRC-Lite occupies a deliberately specialized point in the NVRC family: not a universal operating framework, but a codec optimized for ultra-low decoding complexity.

2. Lightweight representation design

The first major change in NVRC-Lite is architectural. The codec uses HiNeRV as the underlying INR backbone, but it is modified for low complexity by decreasing the number of HiNeRV blocks, reducing channel counts, and replacing a 3D convolutional stem with separate 2D spatial and 1D temporal convolutions. The resulting design is explicitly described as lightweight, yet it incorporates multiple feature grids at different resolutions and feeds those feature grids into multiple blocks at multiple resolutions (Kwan et al., 3 Dec 2025).

The rationale for higher-resolution, multi-scale grids is central. In the low-complexity regime, a small network with fewer blocks and fewer channels has reduced capacity to infer fine detail from coarse latent features alone. NVRC-Lite therefore moves representational burden toward feature grids, allowing the INR to access more localized spatial-temporal detail directly. The paper frames this as a compression-aware trade-off: multi-resolution features improve reconstruction for a fixed compute budget, but they also increase parameter count and therefore bit overhead. NVRC-style end-to-end parameter coding makes that trade-off viable because the grids are themselves compressed efficiently (Kwan et al., 3 Dec 2025).

This representation strategy materially changes the complexity envelope. The paper reports that the resulting multi-scale HiNeRV operates at about $7$ kMACs/pixel, and the complexity table gives NVRC-Lite as $7.8$ kMACs/pixel. By contrast, the original HiNeRV used in NVRC-era settings is described as consuming at least $170$ kMACs/pixel. The architectural claim is therefore not only that NVRC-Lite is smaller, but that higher-resolution grids compensate for that smaller network in a way that preserves useful reconstruction quality at a substantially lower compute budget (Kwan et al., 3 Dec 2025).

The ablation study supports this interpretation. Removing the multi-scale design and reverting to single-scale grids worsens BD-rate, narrows the quality range, and reduces performance across the operating range. The paper’s qualitative conclusion is that multi-scale feature grids are essential for covering a broad rate–quality range at low complexity (Kwan et al., 3 Dec 2025).

3. Hierarchical parameter coding and the octree entropy model

NVRC-Lite inherits the hierarchical coding view of NVRC. The optimization target is still written as

$8$0

where $8$1 is reconstruction distortion, $8$2 is the estimated bitrate, and $8$3 controls the rate–distortion trade-off. In this formulation, $8$4 denotes the INR parameters, $8$5 the entropy-model and quantization parameters for coding $8$6, and $8$7 the higher-level parameters that code $8$8 (Kwan et al., 3 Dec 2025).

The second major modification is the replacement of autoregressive feature-grid entropy coding with an octree-based context model. The paper motivates this change by identifying entropy coding as a practical bottleneck in INR codecs. For a 3D latent grid, an autoregressive model estimates

$8$9

which is compression-effective but requires many sequential decoding steps. For large spatio-temporal grids, this becomes slow and impractical (Kwan et al., 3 Dec 2025).

NVRC-Lite instead partitions the spatio-temporal tensor into $F_\theta$0 blocks and codes only a subset of entries at each step. The conditional model is written as

$F_\theta$1

where $F_\theta$2 is the mask for coding step $F_\theta$3. The paper describes this as an octree-based context model, essentially a 3D extension of checkerboard or quadtree coding. To further reduce sequential depth, it codes two values in every $F_\theta$4 sub-block per step, bringing the number of coding steps down to four (Kwan et al., 3 Dec 2025).

Several accompanying design choices make the model more compression-oriented. The entropy model interleaves coding order across channels, uses a block-wise quantization step size, and introduces a block-wise auxiliary conditional latent learned end-to-end and fed into the entropy model. The paper describes this auxiliary latent as analogous to a hyperprior in classical neural compression. These measures are particularly relevant because the feature grids can be high-dimensional; the text gives an example size of $F_\theta$5. The claimed benefit is that the octree model reduces sequential operations, fits more efficiently within GPU memory constraints, and yields a practical entropy coder for ultra-lightweight INR codecs (Kwan et al., 3 Dec 2025).

4. Optimization procedure and evaluation protocol

The overall compression pipeline follows the NVRC paradigm. For each input video, the codec initializes the INR, optimizes the network parameters $F_\theta$6 for reconstruction fidelity and rate, learns the entropy-model parameters $F_\theta$7 and higher-level coding parameters $F_\theta$8, quantizes the parameters, entropy-codes them, and reconstructs the video by decoding the quantized parameters and evaluating the INR. The distortion term used in the reported NVRC-Lite experiments is MSE (Kwan et al., 3 Dec 2025).

The experimental setup is tightly specified. The paper evaluates on UVG and HEVC-B, using 1080p sequences converted from YUV 4:2:0 to RGB 4:4:4 with BT.601. Each sequence uses 96 frames. Training follows the NVRC optimization pipeline with two stages of 1440 and 120 epochs, a batch size of 4, and MSE as the distortion loss. Evaluation uses PSNR, MS-SSIM, and BD-rate for rate–distortion comparison, while computational practicality is measured with kMACs/pixel, encoding FPS, and decoding FPS (Kwan et al., 3 Dec 2025).

This protocol is consequential for interpreting the reported results. The use of full 1080p sequences and explicit coding-speed measurements indicates that the paper is not only making a rate–distortion claim, but also a deployment-oriented claim about runtime behavior. A plausible implication is that NVRC-Lite is intended to be judged as a codec system rather than solely as a representation model.

5. Reported performance

The primary baseline is C3, which the paper describes as one of the best lightweight INR-based video codecs. Against C3, NVRC-Lite reports the following BD-rate savings: on UVG, $F_\theta$9 in PSNR and $\theta$0 in MS-SSIM; on HEVC-B, $\theta$1 in PSNR and $\theta$2 in MS-SSIM. The paper also reports up to $\theta$3 coding gain in PSNR over x265-medium on UVG and states that it shows strong gains over x265-veryslow as well, though not always in the same direction depending on dataset and metric (Kwan et al., 3 Dec 2025).

The complexity and speed comparison is equally prominent. The table reports C3 at $\theta$4 kMACs/pixel, encoding FPS $\theta$5, and decoding FPS $\theta$6, while NVRC-Lite is reported at $\theta$7 kMACs/pixel, encoding FPS $\theta$8, and decoding FPS $\theta$9 with $\phi$0 in parentheses for the entropy-coding-related comparison. The headline summary is an $\phi$1 encoding speedup and a $\phi$2 decoding speedup (Kwan et al., 3 Dec 2025).

The ablation study attributes these gains to both of the paper’s core design choices. Replacing the octree context model with autoregressive entropy coding makes the entropy coder about $\phi$3 slower and also worsens performance. Removing multi-scale grid features reduces the quality range and hurts rate–distortion performance. The visual examples reported in the paper are consistent with these quantitative findings, showing finer texture, fewer visible artifacts, and better fidelity at lower bitrate relative to C3 (Kwan et al., 3 Dec 2025).

6. Interpretation, limitations, and subsequent developments

NVRC-Lite’s notion of “ultra-lightweight” has two distinct components. On the compute side, it uses a small HiNeRV-based INR with fewer blocks and channels and operates around $\phi$4–$\phi$5 kMACs/pixel. On the bitstream side, it assumes that high-resolution multi-scale features are only worthwhile if they can be compressed efficiently, and the octree context model is introduced precisely to make that entropy coding fast enough to be practical. The codec is therefore lightweight not because it abandons sophisticated parameter coding, but because it adapts that coding to the low-complexity regime (Kwan et al., 3 Dec 2025).

The paper also notes practical caveats. The entropy coder is still unoptimized, and C3’s implementation does not perform actual entropy coding in the same way, which complicates direct speed comparisons. These are not merely incidental details: they delimit how broadly the runtime claims should be generalized. A cautious reading is that the reported speedups establish a strong systems-level advantage within the paper’s implementation setting, while leaving room for further engineering refinement (Kwan et al., 3 Dec 2025).

Within the broader NVRC trajectory, NVRC-Lite can be read as the lightweight specialization of the original NVRC framework. NVRC introduced fully end-to-end, hierarchical compression of INR parameters and reported a $\phi$6 average coding gain over VVC VTM (Random Access) on UVG in PSNR (Kwan et al., 2024). NVRC++ then extended the family in a different direction, proposing a unified architecture with four fixed complexity levels from $\phi$7 kMACs/pixel to $\phi$8 kMACs/pixel and reporting decoding speed up to $\phi$9 faster than NVRC while delivering comparable performance (Kwan et al., 26 Jun 2026). This suggests that NVRC-Lite’s historical significance lies in demonstrating that the NVRC philosophy can remain effective even when pushed into an ultra-lightweight operating regime, provided that both the representation and the entropy model are redesigned accordingly.