Research map / Label-efficient and robust learning / Domain adaptation and generalization
Multirate Neural Image Compression with Adaptive Lattice Vector Quantization
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 adaptive lattice vector quantization for learned image compression that scales the lattice basis for variable-rate coding and learns an invertible linear transform of the basis for domain adaptation.
Problem
Existing lattice vector quantization (LVQ) based neural image codecs use a fixed predefined lattice basis, so they need separate models per rate-distortion trade-off and cannot adapt to new source domains.
Approach
Control bit rate by scaling the diamond lattice basis (changing lattice point density) in a single network, and perform domain adaptation by learning an invertible linear transform (via matrix exponential rotation and diagonal scaling) of the fixed coset-representable lattice basis while the network is frozen.
Key results
- LVQ variable-rate Cheng2020 yields BD-Rate of -9.56% (Kodak) vs SQ-s baseline and about -19% vs SQ variable-rate on Tecnick style comparisons; SQ variable-rate Cheng2020 shows 29.20% BD-rate drop from fixed to variable rate on Tecnick versus 4.37% for LVQ.
- Domain adaptation of the lattice basis on screen content (SCI1K) gives BD-Rate -0.99% and BD-PSNR 0.052 dB; PACS cartoon gains range from -0.47% to -1.59%.
- The learned linear mapping has only about 0.06% of total network parameters, and single-network LVQ variable-rate models keep roughly 80% of fixed-rate LVQ gains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SQ-based variable-rate quantization with adaptive LVQ | Cheng2020 · Kodak | BD-Rate: -6.82% | Helped |
| replace SQ-based variable-rate quantization with adaptive LVQ | Cheng2020 · CLIC valid | BD-Rate: -15.63% | Helped |
| replace SQ-based variable-rate quantization with adaptive LVQ | Cheng2020 · Tecnick | BD-Rate: -17.39% | Helped |
| replace SQ-based variable-rate quantization with adaptive LVQ | Cheng2020 checkerboard · Kodak | BD-Rate: -10.06% | Helped |
| replace SQ-based variable-rate quantization with adaptive LVQ | Cheng2020 checkerboard · CLIC valid | BD-Rate: -8.45% | Helped |
| replace SQ-based variable-rate quantization with adaptive LVQ | Cheng2020 checkerboard · Tecnick | BD-Rate: -7.85% | Helped |
| replace SQ-based variable-rate quantization with adaptive LVQ | MBT2018 mean · Kodak | BD-Rate: -3.92% | Helped |
| replace SQ-based variable-rate quantization with adaptive LVQ | MBT2018 mean · CLIC valid | BD-Rate: -6.62% | Helped |
| replace SQ-based variable-rate quantization with adaptive LVQ | MBT2018 mean · Tecnick | BD-Rate: -8.44% | Helped |
| adapt the lattice basis to the target domain using a learned linear mapping | Cheng2020 LVQ · SCI1K | BD-Rate: -0.99% | Helped |
| adapt the lattice basis to the target domain using a learned linear mapping | Cheng2020 LVQ · PACS cartoon | BD-Rate: -0.47% to -1.59% | Helped |
Surprising findings
- SQ-based variable-rate models lose much more R-D performance than LVQ variable-rate models relative to fixed-rate training (e.g., 29.20% vs 4.37% BD-rate drop on Tecnick for Cheng2020).
Limitations
- Modest R-D gains in the target domain from lattice basis optimization alone, since only the linear mapping is updated.
- Relies on lattices with coset representations for fast encoding, restricting the basis to the transformed diamond lattice.
Experimental setup
Datasets: Flickr2W SCI1K PACS
Benchmarks: Kodak CLIC validation Tecnick SCI1K PACS cartoon
Metrics: BD-Rate BD-PSNR PSNR bpp
Baselines: Cheng2020 SQ Cheng2020 checkerboard SQ MBT2018 mean SQ LVQAC LVQ fixed-rate models
Builds on: LVQAC Cheng2020 MBT2018 lattice basis reduction
Closest papers by meaning
- Learning Optimal Lattice Vector Quantizers for End-to-end Neural Image Compression (NeurIPS 2024 · similarity 0.94)
- Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding (ICLR 2025 · similarity 0.91)
- Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression (NeurIPS 2025 · similarity 0.87)
- HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression (NeurIPS 2026 · similarity 0.86)
- On Quantizing Neural Representation for Variable-Rate Video Coding (ICLR 2025 · similarity 0.86)
- Binary Quadratic Quantization: Beyond First-Order Quantization for Real-Valued Matrix Compression (NeurIPS 2025 · similarity 0.85)
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.