Research map / Image generation / Image restoration
Correcting Quantization-Induced Gradient Mismatch in Neural Image Compression
Changhao Peng, Yuqi Ye, Wei Gao
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 |
|---|---|---|---|
| add MSE-guided gradients on mu when |y-mu|<1/2 (ablation) | Checkerboard · Kodak | BD-BR (%): 3.29 -> 1.79 | Improved |
| add gradients transfer (k=0.01) (ablation) | Checkerboard · Kodak | BD-BR (%): 3.29 -> 2.32 | Improved |
| combine MSE-guided gradients and gradients transfer (ablation) | Checkerboard · Kodak | BD-BR (%): 3.29 -> 1.20 | Improved |
| combine both gradient corrections | ELIC · Kodak | BD-BR (%): -8.00 -> -9.55 | Improved |
| combine both gradient corrections | MLIC++ · Kodak | BD-BR (%): -15.26 -> -16.56 | Improved |
| combine both gradient corrections | DCAE · Kodak | BD-BR (%): -17.36 -> -18.16 | Improved |
| fine-tune pretrained model without improvement (ablation) | Checkerboard · Kodak | BD-BR (%): 3.29 -> 3.29 | No clear effect |
| apply MSE-guided gradients also when round(y-mu)!=0 (ablation) | Checkerboard MSE-Grad · Kodak | BD-BR (%): 1.79 -> 1.84 | Hurt |
| lower gradient-transfer k from 0.01 to 0.001 (ablation) | Checkerboard GT · Kodak | BD-BR (%): 2.32 -> 2.85 | Hurt |
| raise gradient-transfer k to 0.1 (ablation) | Checkerboard GT · Kodak | BD-BR (%): NA (training NaN) | Hurt |
Closest papers by meaning
- Variational image compression with a scale hyperprior (ICLR 2018 · similarity 0.89)
- Soft then Hard: Rethinking the Quantization in Neural Image Compression (ICML 2021 · similarity 0.89)
- Towards Efficient Image Compression Without Autoregressive Models (NeurIPS 2023 · similarity 0.89)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.88)
- Diversify, Contextualize, and Adapt: Efficient Entropy Modeling for Neural Image Codec (NeurIPS 2024 · similarity 0.87)
- Causal Context Adjustment Loss for Learned Image Compression (NeurIPS 2024 · 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.
