Research map / 3D vision / NeRF and radiance fields
How Far Can We Compress Instant-NGP-Based NeRF?
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.
Introduces Context-based NeRF Compression (CNC), using level-wise and dimension-wise context models plus hash-fusion with occupancy grids to entropy-code binarized Instant-NGP hash embeddings.
Problem
Hybrid NeRF representations such as Instant-NGP require large storage due to learnable feature embeddings, which limits deployment of NeRFs at scale.
Approach
Model the entropy of binarized multi-resolution hash embeddings with context models that exploit cross-level (coarse-to-fine) and cross-dimension (3D voxel to 2D plane projections) dependencies, and use hash collision handling via occupancy-grid-weighted AOE to improve probability estimates.
Key results
- Achieves 86.7% (Synthetic-NeRF) and 82.3% (Tanks and Temples) BD-rate reduction against the SOTA NeRF compression method BiRF.
- Achieves over 100x size reduction on Synthetic-NeRF and 70x on Tanks and Temples versus Instant-NGP while improving fidelity.
- Disabling context models raises BD-rate by 30.8% to 43.7% on Synthetic-NeRF, showing context models are the main source of gains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply CNC compression to Instant-NGP | Instant-NGP · Synthetic-NeRF | model size: 100⇥ size reduction | Helped |
| apply CNC compression to Instant-NGP | Instant-NGP · Tanks and Temples | model size: 70⇥ size reduction | Helped |
| use CNC instead of BiRF | Instant-NGP-based NeRF · Synthetic-NeRF | BD-rate: 86.7% | Helped |
| use CNC instead of BiRF | Instant-NGP-based NeRF · Tanks and Temples | BD-rate: 82.3% | Helped |
| disable 2D and 3D context models (ablation) | CNC · Synthetic-NeRF | BD-rate: +43.7%/+78.8% | Hurt |
| add dimension-wise context (ablation) | CNC · Synthetic-NeRF | BD-rate: +30.8%/+54.3% without it | Helped |
| add 3D context (ablation) | CNC · Synthetic-NeRF | BD-rate: +5.7%/+9.2% without it | Helped |
| reverse level-wise context order to fine-to-coarse (ablation) | CNC · Synthetic-NeRF | BD-rate: +33.8%/+53.0% | Hurt |
| retain invalid feature vectors instead of discarding them (ablation) | CNC · Synthetic-NeRF | BD-rate: +43.7%/+78.8% | Hurt |
| discard vectors judged invalid using AOE (ablation) | CNC · Synthetic-NeRF | BD-rate: +32.6%/+59.1% | Helped |
| increase the number of context levels beyond 3 (ablation) | CNC · Synthetic-NeRF | rate-distortion performance | Mixed |
| increase feature dimension F (ablation) | CNC · Synthetic-NeRF | rate-distortion performance: BD-rate: +13.0%; BD-rate: +16.6%; BD-rate: +26.4% | Helped |
Surprising findings
- On Tanks and Temples, the compressed model surpasses its own fidelity upper bound, which the authors attribute to the entropy constraint acting as regularization against overfitting.
- Over-discarding hash embeddings based on occupancy alone collapses fidelity to about 27.2 dB PSNR, so precise AOE-based discarding is needed.
Limitations
- Training is about 1.3x slower than the model without context models.
- Evaluated only on static scenes (Synthetic-NeRF and Tanks and Temples).
Experimental setup
Datasets: Synthetic-NeRF Tanks and Temples
Benchmarks: Synthetic-NeRF Tanks and Temples
Metrics: PSNR SSIM LPIPS BD-rate model size (MB)
Baselines: BiRF MaskDWT VQRF Re:NeRF DVGO Plenoxels TensoRF CCNeRF Instant-NGP K-Planes
Builds on: Instant-NGP BiRF NerfAcc
Closest papers by meaning
- Neural NeRF Compression (ICML 2024 · similarity 0.91)
- NeRFCodec: Neural Feature Compression Meets Neural Radiance Fields for Memory-Efficient Scene Representation (CVPR 2024 · similarity 0.89)
- CodecNeRF: Toward Fast Encoding and Decoding, Compact, and High-quality Novel-view Synthesis (AAAI 2025 · similarity 0.88)
- A General Implicit Framework for Fast NeRF Composition and Rendering (AAAI 2024 · similarity 0.86)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.86)
- Ced-NeRF: A Compact and Efficient Method for Dynamic Neural Radiance Fields (AAAI 2024 · similarity 0.86)
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.