Research map / Image generation / GANs
Generative Low-bitwidth Data Free Quantization
Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, Mingkui Tan
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 CE loss to generator training (without BNS loss) (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 30.70 -> 54.51 | Improved |
| add BNS loss to generator training (without CE loss) (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 30.70 -> 44.40 | Improved |
| add BNS loss on top of CE loss for generator (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 54.51 -> 63.91 | Improved |
| add CE loss on top of BNS loss for generator (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 44.40 -> 63.91 | Improved |
| add KD loss on top of CE loss for quantized model (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 55.55 -> 63.91 | Improved |
| add CE loss on top of KD loss for quantized model (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 62.98 -> 63.91 | Improved |
| add fixed batch normalization statistics (FBNS) (ablation) | GDFQ · CIFAR-10 ResNet-20 W4A4 | top-1 accuracy: 89.21 -> 90.23 | Improved |
| add fixed batch normalization statistics (FBNS) (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 61.12 -> 63.91 | Improved |
| alternate G and Q training instead of separate training (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 61.81 -> 63.91 | Improved |
| never stop generator training instead of stopping at threshold 99.50% (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 59.67 -> 63.91 | Improved |
| raise generator stopping threshold from 90.00% to 99.50% (ablation) | GDFQ · CIFAR-100 ResNet-20 W4A4 | top-1 accuracy: 57.20 -> 59.67 | Improved |
| use fake data GDFQ instead of ZeroQ | GDFQ · ImageNet ResNet-18 W4A4 | top-1 accuracy: 26.04 -> 60.60 | Improved |
Closest papers by meaning
- Enhancing Diversity for Data-free Quantization (CVPR 2025 · similarity 0.87)
- Zero-Shot Adversarial Quantization (CVPR 2021 · similarity 0.87)
- Diverse Sample Generation: Pushing the Limit of Generative Data-Free Quantization (TPAMI 2023 · similarity 0.87)
- Sharpness-Aware Data Generation for Zero-shot Quantization (ICML 2024 · similarity 0.85)
- Style Quantization for Data-Efficient GAN Training (CVPR 2025 · similarity 0.85)
- Quantization without Tears (CVPR 2025 · similarity 0.84)
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.
