Research map / Probabilistic and scientific modeling / Variational and generative sampling
Multi-Sample Training for Neural Image Compression
Tongda Xu, Yan Wang, Dailan He, Chenjian Gao, Han Gao, Kunzan Liu, Hongwei Qin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with multiple-sample MIX objective (sample size 8) instead of single-sample | MS-NIC-MIX · Kodak model A | PSNR BD-BR (%): -3.847 | Improved |
| train with multiple-sample DMS objective (sample size 16) instead of single-sample | MS-NIC-DMS · Kodak model A | PSNR BD-BR (%): -4.929 | Improved |
| apply MS-NIC to autoregressive-context model | MS-NIC-DMS · Kodak model B | PSNR BD-BR (%): -2.378 | Improved |
| apply MS-NIC to autoregressive-context model | MS-NIC-MIX · Kodak model B | MS-SSIM BD-BR (%): 2.238 | Hurt |
| use original IWAE objective without direct-y trick | IWAE · Kodak model A | PSNR BD-BR (%): 64.23 | Hurt |
| use IWAE-DReG gradient instead of pathwise gradient (ablation) | MS-NIC-MIX (5 samples) · Kodak lambda=0.015 | R-D cost: 1.003 -> 1.058 | Hurt |
| remove direct-y trick (feed noisy y~ to z inference) (ablation) | 2-level VAE · Kodak lambda=0.015 | R-D cost: 1.017 -> 1.493 | Hurt |
| increase batch size x3 to x16 instead of importance-weighted samples (ablation) | Baseline · Kodak lambda=0.015 | R-D cost: 1.017 -> 1.013 | No clear effect |
| use MS-NIC-MIX sample size 8 vs 16 (ablation) | MS-NIC-MIX · Kodak lambda=0.015 | R-D cost: 0.9988 -> 1.013 | Mixed |
| increase MS-NIC-DMS sample size 3,3 to 16,16 (ablation) | MS-NIC-DMS · Kodak lambda=0.015 | R-D cost: 1.010 -> 0.9954 | Improved |
| increase IWAE sample size without direct-y (ablation) | IWAE · Kodak lambda=0.015 | R-D cost: 1.400 -> 1.748 | Hurt |
| use universal quantization instead of deterministic rounding (ablation) | Universal Quantization · Kodak image 1 | R-D cost: 0.7552 -> 0.9080 | Hurt |
Closest papers by meaning
- Neural Image Compression: Generalization, Robustness, and Spectral Biases (NeurIPS 2023 · similarity 0.88)
- Improving Inference for Neural Image Compression (NeurIPS 2020 · similarity 0.87)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.87)
- Soft then Hard: Rethinking the Quantization in Neural Image Compression (ICML 2021 · similarity 0.87)
- C3: High-Performance and Low-Complexity Neural Compression from a Single Image or Video (CVPR 2024 · similarity 0.85)
- Neural Video Compression with Feature Modulation (CVPR 2024 · 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.
