Research map / Architectures and efficiency / Quantization
GenQ: Quantization in Low Data Regimes with Generative Synthetic Data
Yuhang Li, Youngeun Kim, Donghyun Lee, Souvik Kundu, Priyadarshini Panda
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 |
|---|---|---|---|
| use GenQ text-to-image synthetic data instead of Genie-D inversion data (MobileNet-b W2A4) | BRECQ PTQ · ImageNet data-free PTQ W2A4 | top-1 accuracy (%): 38.01 -> 47.31 | Improved |
| use GenQ data instead of Genie-D on ResNet architectures | BRECQ PTQ · ImageNet data-free PTQ W4A4 | top-1 accuracy (%): 75.35 -> 75.47 (ResNet-50) | No clear effect |
| use GenQ synthetic data instead of TexQ GAN data for W3A3 QAT | ResNet-50 QAT · ImageNet data-free QAT W3A3 | top-1 accuracy (%): 25.27 -> 73.99 | Improved |
| add GenQ data-scarce synthetic data to 1k real images | ResNet-50 W3A3 QAT · ImageNet 1-shot | top-1 accuracy (%): 68.88% with 1k real, similar to QAT with 300x more data | Improved |
| increase number of synthetic images for PTQ (ablation) | GenQ · ResNet-50 W2A4 PTQ | top-1 accuracy (%) | Improved |
| increase energy-score filtering ratio (ablation) | GenQ · ResNet-18 W2A4 PTQ | top-1 accuracy (%) | Improved |
| increase BN-sensitivity filtering ratio beyond 50% (ablation) | GenQ · ResNet-18 W2A4 PTQ | top-1 accuracy (%) | Hurt |
| use text-to-image synthesis instead of image inversion | synthetic data generation · 1000 ImageNet images | latency: 152 min -> 24 min | Improved |
| use GenQ data synthesized on a different model for quantization (transfer) | GenQ · cross-model PTQ W2A4 | top-1 accuracy (%): drop of 0.2-0.5% vs 0.6-33% for existing methods | No clear effect |
Later papers that compared with this method
1 later accepted paper reports GenQ as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Accurate Zero-shot Quantization via Hierarchical Teacher-Assistant Distillation (ECCV 2026 · better on 6 of 6 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TexQ — TexQ: Zero-shot Network Quantization with Texture Feature Distribution Calibration (NeurIPS 2023) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Efficient Generative Modeling with Residual Vector Quantization-Based Tokens (ICML 2025 · similarity 0.88)
- Enhancing Diversity for Data-free Quantization (CVPR 2025 · similarity 0.86)
- SqueezeLLM: Dense-and-Sparse Quantization (ICML 2024 · similarity 0.86)
- SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer (CVPR 2025 · similarity 0.86)
- Rethinking the Objectives of Vector-Quantized Tokenizers for Image Synthesis (CVPR 2024 · similarity 0.86)
- Robustness-Guided Image Synthesis for Data-Free Quantization (AAAI 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.
