Research map / Architectures and efficiency / RNNs and normalization
Neural Block Compression: Variable Bitrates Feature Blocks for Texture Representation
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 Neural Block Compression (NBC), a multi-resolution neural feature-block texture codec with variable quantization bitrates per mip-level and residual fusion, reaching very low BPPC for material textures.
Problem
Traditional block-based texture compression (BC/ASTC) and neural texture compression (NTC) perform poorly at low bitrates (around 0.1-0.2 bits per pixel per channel) because they operate in color space or use a fixed quantization bitrate.
Approach
Extends block compression from color space to a neural feature domain: a hierarchy of multi-resolution quantized feature grids, with higher bitrates for coarser feature mip-levels (8-4-2-1 bits) and a residual fusion that adds finer-level low-bit displacements to interpolated coarser features.
Key results
- At ~0.1 BPPC, NBC's PSNR exceeds JPEG XL by about 13% and outweighs AVIF and JPEG XL at 0.2 BPPC.
- NBC reaches ~0.05 BPPC with PSNR 31.85 dB, where BC and ASTC cannot reach 0.2 BPPC.
- In the iso-quality comparison at ~37 dB PSNR, NBC uses under one-third of BC's storage and 0.7 MB less than NTC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NBC's variable-bitrate neural feature blocks | Texture compression · 16-texture dataset at low bitrates | PSNR | Helped |
| Use NBC instead of BC at comparable quality | Texture compression · Texture set, approximately 37 dB PSNR | storage: 60 MB vs 17.5 MB | Helped |
| Use NBC instead of NTC at comparable quality | Texture compression · Texture set, approximately 37 dB PSNR | storage: 18.2 MB vs 17.5 MB | Helped |
| Use full-precision feature quantization instead of 8-4-2-1 (ablation) | NBC · Texture set | PSNR: 33.02 -> 35.09 | Helped |
| Use full-precision feature quantization instead of 8-4-2-1 (ablation) | NBC · Texture set | BPPC: 0.101 -> 2.252 | Hurt |
| Vary bitrate configuration from 8-4-2-1 to 8-4-4-4 (ablation) | NBC · Texture set | PSNR: 33.02 -> 34.78 | Helped |
| Vary bitrate configuration from 8-4-2-1 to 8-4-4-2 (ablation) | NBC · Texture set | PSNR: 33.02 -> 33.86 | Helped |
| Vary bitrate configuration from 8-4-2-1 to 8-4-2-2 (ablation) | NBC · Texture set | PSNR: 33.02 -> 33.19 | Helped |
| Vary bitrate configuration from 8-4-2-1 to 2-2-2-2 (ablation) | NBC · Texture set | PSNR: 33.02 -> 32.36 | Hurt |
| Remove feature reuse (ablation) | NBC · Texture set, 2-bit quantization | PSNR: 32.36 -> 32.78 | Helped |
| Remove feature reuse (ablation) | NBC · Texture set, 2-bit quantization | BPPC: 0.141 -> 0.176 | Hurt |
Surprising findings
- Full-precision configuration is only 2.07 dB better than the default 8-4-2-1 configuration but uses 22x more bits.
- Removing feature reuse gains only 0.42 dB PSNR at 25% more bits.
Limitations
- NTC had to be reimplemented because source code was unavailable, so the NTC comparison depends on the reimplementation.
- At high bitrates, the paper concedes that JPEG XL is marginally better on diffuse textures than NBC.
- Compression ratios for the baselines are only approximately matched to target BPPC.
Experimental setup
Datasets: ambientCG EISKO KaiMoisch PolyHaven textures (16 texture sets)
Benchmarks: 16-texture-set dataset from ambientCG, EISKO, KaiMoisch, PolyHaven
Metrics: PSNR SSIM LPIPS FLIP BPPC
Baselines: NTC BC ASTC AVIF JPEG XL MLIC++
Builds on: Neural Texture Compression (NTC) Block Compression (BC1) auto-decoder (DeepSDF-style) simulated quantization (Ballé et al.)
Closest papers by meaning
- NeRFCodec: Neural Feature Compression Meets Neural Radiance Fields for Memory-Efficient Scene Representation (CVPR 2024 · similarity 0.83)
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.83)
- Paint-it: Text-to-Texture Synthesis via Deep Convolutional Texture Map Optimization and Physically-Based Rendering (CVPR 2024 · similarity 0.82)
- Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding (ICLR 2025 · similarity 0.82)
- Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation (ICLR 2025 · similarity 0.82)
- Decouple Distortion from Perception: Region Adaptive Diffusion for Extreme-low Bitrate Perception Image Compression (CVPR 2025 · similarity 0.82)
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.