Research map / Architectures and efficiency / Quantization
Qimera: Data-free Quantization with Synthetic Boundary Supporting Samples
Kanghyun Choi, Deokki Hong, Noseong Park, Youngsok Kim, Jinho Lee
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 superposed embeddings (SE) (ablation) | GDFQ · ImageNet ResNet-50 4w4a | top-1 accuracy: 52.12 -> 64.09 | Improved |
| add superposed embeddings (SE) (ablation) | GDFQ · CIFAR-100 ResNet-20 4w4a | top-1 accuracy: 63.39 -> 64.55 | Improved |
| add disentanglement mapping (DM) to SE (ablation) | GDFQ + SE · ImageNet ResNet-50 4w4a | top-1 accuracy: 64.09 -> 66.06 | Improved |
| add disentanglement mapping (DM) to SE (ablation) | GDFQ + SE · CIFAR-100 ResNet-20 4w4a | top-1 accuracy: 64.55 -> 64.76 | No clear effect |
| add extracted embedding initialization (EEI) to SE (ablation) | GDFQ + SE · CIFAR-100 ResNet-20 4w4a | top-1 accuracy: 64.55 -> 65.23 | Improved |
| add DM and EEI together to SE (ablation) | GDFQ + SE · ImageNet ResNet-50 4w4a | top-1 accuracy: 66.06 -> 66.25 | Improved |
| freeze embeddings after extracted initialization (no training) (ablation) | Qimera · ImageNet ResNet-50 4w4a | top-1 accuracy: 66.25 -> 63.20 | Hurt |
| apply Mixup-style input mixing on GDFQ samples | GDFQ · ImageNet ResNet-50 4w4a | top-1 accuracy: 52.12 -> 59.25 | Improved |
| apply Cutmix on GDFQ samples | GDFQ · CIFAR-100 ResNet-20 4w4a | top-1 accuracy: 63.39 -> 62.74 | Hurt |
| lower generator noise sigma_z from 1.0 to 0.25 (ablation) | GDFQ · ImageNet ResNet-50 4w4a | top-1 accuracy: 52.12 -> 55.30 | Improved |
| raise generator noise sigma_z from 1.0 to 2.0 (ablation) | GDFQ · ImageNet ResNet-50 4w4a | top-1 accuracy: 52.12 -> 46.14 | Hurt |
| use 8 DM layers instead of 1 (ablation) | Qimera · CIFAR-100 ResNet-20 4w4a | top-1 accuracy: 65.10 -> 64.11 | Hurt |
Later papers that compared with this method
1 later accepted paper reports Qimera as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- TexQ: Zero-shot Network Quantization with Texture Feature Distribution Calibration (NeurIPS 2023 · better on 1 of 1 shared result)
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
- Enhancing Diversity for Data-free Quantization (CVPR 2025 · similarity 0.86)
- Quantization without Tears (CVPR 2025 · similarity 0.85)
- Data-Free Quantization via Pseudo-label Filtering (CVPR 2024 · similarity 0.85)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.85)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.85)
- QuantDemoire: Quantization with Outlier Aware for Image Demoiréing (NeurIPS 2026 · 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.
