Research map / Image generation / Autoregressive and latent generation
Robustness-Guided Image Synthesis for Data-Free Quantization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Robustness-Guided Image Synthesis (RIS), which adds feature- and prediction-level robustness losses under input/weight perturbations plus soft-label diversity to a generator, improving synthetic-image quality and data-free quantization accuracy.
Problem
Data-free quantization synthesizes images with classification and BN losses, but the synthetic images are low-semantic and homogeneous, limiting downstream compressed-model accuracy.
Approach
Treat the teacher's feature and prediction inconsistency under input augmentations and weight perturbations as an image robustness measure, and use it as a generator loss; additionally replace one-hot labels with multiple soft labels having minimal pairwise correlation to increase diversity.
Key results
- On ImageNet ResNet-50 W4A4, RIS improves GDFQ from 52.12% to 62.16% (+10.04) and DSG from 54.68% to 63.40% (+8.72).
- With AIT + ARC on ResNet-50 W4A4, RIS raises top-1 from 68.27% to 71.54%.
- On CIFAR-100 synthetic images, RIS reaches FID 87.48 and IS 11.81 versus GDFQ's 142.42 and 1.95.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add RIS to GDFQ, DSG, ARC, or AIT | Data-free quantization methods · CIFAR-10/100 and ImageNet, W4A4/W5A5 | top-1 accuracy: +0.79; +0.21; +2.11; +2.96; +2.21; +1.37; +10.04; +3.35; +1.54; -0.06; +2.57; +2.30; +3.01; +0.31; +8.72; +2.93; +2.89; | Mixed |
| Use RIS synthetic images instead of baseline-generated images | Synthetic image generation · CIFAR-10 | IS: 3.96 -> 10.4 | Helped |
| Use RIS synthetic images instead of baseline-generated images | Synthetic image generation · CIFAR-10 | FID: 120.91 -> 97.45 | Helped |
| Use RIS synthetic images instead of baseline-generated images | Synthetic image generation · CIFAR-100 | IS: 1.95 -> 11.81 | Helped |
| Use RIS synthetic images instead of baseline-generated images | Synthetic image generation · CIFAR-100 | FID: 142.42 -> 87.48 | Helped |
| Add input perturbation (ablation) | RIS quantization on CIFAR-100 ResNet-20 W4A4 · CIFAR-100 | top-1 accuracy: 63.39 -> 64.38 | Helped |
| Add weight perturbation (ablation) | RIS quantization on CIFAR-100 ResNet-20 W4A4 · CIFAR-100 | top-1 accuracy: 63.39 -> 64.48 | Helped |
| Add soft labels (ablation) | RIS quantization on CIFAR-100 ResNet-20 W4A4 · CIFAR-100 | top-1 accuracy: 63.39 -> 63.99 | Helped |
| Combine input perturbation, weight perturbation, and soft labels (ablation) | RIS quantization on CIFAR-100 ResNet-20 W4A4 · CIFAR-100 | top-1 accuracy: 63.39 -> 65.50 | Helped |
| Use Gaussian noise, adversarial perturbation, or dropout for weight perturbation (ablation) | RIS quantization on CIFAR-100 ResNet-20 W4A4 · CIFAR-100 | top-1 accuracy: 63.39 -> 64.48; 63.39 -> 63.65; 63.39 -> 64.14 | Mixed |
| Vary input augmentation strategy among noise, translation, resize, and random selection (ablation) | RIS quantization on CIFAR-100 ResNet-20 W4A4 · CIFAR-100 | top-1 accuracy: 63.39 -> 64.38; 63.39 -> 64.42; 63.39 -> 64.40; 63.39 -> 64.88 | Helped |
| Set ε to 0, 0.05, 0.2, or 0.5 instead of 0.1 (ablation) | RIS quantization on CIFAR-100 ResNet-20 W4A4 · CIFAR-100 | top-1 accuracy: 65.50 -> 64.95; 65.50 -> 65.05; 65.50 -> 64.89; 65.50 -> 64.08 | Mixed |
Surprising findings
- Pure Gaussian noise inputs achieve performance similar to synthesized images, which the authors use to justify threshold calibration.
- DSG with RIS slightly degrades on CIFAR-10 W5A5 (-0.06%), attributed to near-teacher accuracy.
Limitations
- Generalization to methods without a generator is unexplored.
- Handling large numbers of out-of-distribution data in the wild is left for future work.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet
Benchmarks: CIFAR-10 CIFAR-100 ImageNet
Metrics: top-1 accuracy top-5 accuracy FID IS
Baselines: GDFQ DSG AutoReCon (ARC) AIT Qimera
Builds on: GDFQ DSG AutoReCon AIT Qimera
Closest papers by meaning
- Data-Free Quantization via Pseudo-label Filtering (CVPR 2024 · similarity 0.86)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.86)
- Enhancing Diversity for Data-free Quantization (CVPR 2025 · similarity 0.86)
- Rethinking the Objectives of Vector-Quantized Tokenizers for Image Synthesis (CVPR 2024 · similarity 0.86)
- Bridging the Gap between Gaussian Diffusion Models and Universal Quantization for Image Compression (CVPR 2025 · similarity 0.85)
- HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression (NeurIPS 2026 · 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.