Research map / Architectures and efficiency / Quantization
On Quantizing Neural Representation for Variable-Rate Video Coding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeuroQuant, a post-training quantization method that achieves variable-rate coding for per-video implicit neural representation codecs (INR-VC) by adjusting quantization parameters of pretrained weights instead of retraining per bitrate.
Problem
Non-generalized INR-based video codecs need separate weight training for each target bitrate, which is very time-consuming, and existing mixed-precision and PTQ methods assume inter-layer independence that does not hold for INR-VC.
Approach
Frames variable-rate INR-VC as mixed-precision bit allocation with a Hessian-vector-product sensitivity criterion that accounts for off-diagonal Hessian terms and perturbation direction, then calibrates quantization steps and rounding variables network-wise with channel-wise quantization using an MSE-oriented Gauss-Newton objective.
Key results
- At INT2 on UVG, NeuroQuant reaches 28.37 dB average PSNR vs 26.17 dB for the best baseline (QDrop) and 22.93 dB for FFNeRV QAT.
- At INT4, NeuroQuant averages 32.08 dB vs 31.75 dB for HiNeRV QAT; encoding time for a new bitrate drops from 22.2 h to 2.8 h for HiNeRV (7.9x speedup).
- Versus NeRV and HNeRV with 8-bit quantization, NeuroQuant gives compression gains of 27.8% and 25.5%, and a further 4.8% over HiNeRV.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeuroQuant instead of the strongest listed baseline, HiNeRV | 6-bit INR-VC on UVG · UVG average | PSNR (dB): 32.71 -> 32.84 | Helped |
| Use NeuroQuant instead of the strongest listed baseline, HiNeRV | 4-bit INR-VC on UVG · UVG average | PSNR (dB): 31.75 -> 32.08 | Helped |
| Use NeuroQuant instead of the strongest listed baseline, QDrop | 2-bit INR-VC on UVG · UVG average | PSNR (dB): 26.17 -> 28.37 | Helped |
| Use NeuroQuant instead of DeepCABAC | NeRV variable-rate coding · Beauty, Jockey, ReadyS average at bpp 0.013 | PSNR (dB): 28.86 -> 29.37 | Helped |
| Adjust quantization parameters on a pretrained model rather than retraining for a new bitrate | NeRV, HNeRV, and HiNeRV · Support a new bitrate | Encoding time: 1.8 h -> 0.4 h (×4.5); 4.7 h -> 1.0 h (×4.7); 22.2 h -> 2.8 h (×7.9) | Helped |
| Increase calibration batch size from 1 to 8 (ablation) | NeuroQuant · NeRV-3M, Beauty/Jockey/ReadyS | PSNR (dB): 29.02 -> 29.09 | Helped |
| Vary calibration batch size from 1 to 8 (ablation) | NeuroQuant · NeRV-3M, Beauty/Jockey/ReadyS | PSNR (dB): 29.02 -> 29.09 | Helped |
| Set learning rate to 3e-3 instead of 5e-3 (ablation) | NeuroQuant · NeRV-3M, Beauty/Jockey/ReadyS | PSNR (dB): 29.04 -> 29.12 | Helped |
| Vary regularization weight λ across 0.1, 0.01, and 0.001 (ablation) | NeuroQuant calibration objective · NeRV-3M, Beauty/Jockey/ReadyS | PSNR (dB): 29.06; 29.04; 29.03 | No clear effect |
| Use channel-wise rather than layer-wise quantization with NeuroQuant (ablation) | NeuroQuant · NeRV-3M, Beauty/Jockey/ReadyS | PSNR (dB): 27.90 -> 29.06 | Helped |
Surprising findings
- Higher inter-layer dependence in INR-VC (from longer overfitting) increases quantization-induced loss change yet yields better representation performance.
- A higher Hessian-sensitive perturbation can produce a smaller loss increase, so perturbation direction matters beyond magnitude.
Limitations
- Variable-rate range is bounded by available bitwidths; pruning suggested as remedy.
- Not a fully joint rate-distortion optimization; uses simple lossless entropy coding rather than CABAC or learned context models.
- Quantization below INT3 shows noticeable degradation, so INT3 and below are not recommended for variable-rate use.
Experimental setup
Datasets: UVG
Benchmarks: UVG
Metrics: PSNR bpp BD-Rate encoding time
Baselines: HNeRV naive PTQ FFNeRV QAT HiNeRV QAT AdaRound BRECQ QDrop RDOPTQ DeepCABAC Gomes et al. DCVC DCVC-DC x264
Builds on: AdaRound BRECQ HAWQ-V2 HNeRV QDrop
Closest papers by meaning
- Neural Video Compression with Feature Modulation (CVPR 2024 · similarity 0.89)
- NVRC: Neural Video Representation Compression (NeurIPS 2024 · similarity 0.89)
- PNVC: Towards Practical INR-based Video Compression (AAAI 2025 · similarity 0.88)
- Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding (ICLR 2025 · similarity 0.87)
- Combining Frame and GOP Embeddings for Neural Video Representation (CVPR 2024 · similarity 0.87)
- Towards Practical Real-Time Neural Video Compression (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.