Research map / Architectures and efficiency / Pruning and efficient tuning
Scaling Backwards: Minimal Synthetic Pre-training?
Ryo Nakamura, Ryu Tadokoro, Ryosuke Yamada, Yuki M Asano, Iro Laina, Christian Rupprecht, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka
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 |
|---|---|---|---|
| pre-train on a single fractal image with LPCE loss instead of 1M-image FractalDB | ViT-Tiny · CIFAR-100 fine-tuning | top-1 accuracy: 81.6 -> 84.2 | Improved |
| pre-train on ImageNet-21k scale versus single fractal 1p-frac for ViT-B | ViT-B · ImageNet-1k fine-tuning | top-1 accuracy: 81.8 -> 82.1 | Improved |
| use single Gaussian or uniform noise image with LPCE instead of a fractal (ablation) | ViT-Tiny · CIFAR-100 fine-tuning | top-1 accuracy: 84.2 -> 1.1 | Hurt |
| reduce perturbation degree Delta from 0.05 to 0.01 (ablation) | 1p-frac · CIFAR-100 fine-tuning | top-1 accuracy: 83.0 -> 19.9 | Hurt |
| increase perturbation degree beyond 0.1 (ablation) | 1p-frac · CIFAR-100 fine-tuning | top-1 accuracy: 84.2 -> 82.6 | Hurt |
| use the most complex fractal shape (sigma 3.5) instead of simpler ones (ablation) | 1p-frac · CIFAR-100 fine-tuning | top-1 accuracy: 81.3 -> 84.2 | Improved |
| increase numerical integration sampling points L from 16 to 1000 (ablation) | 1p-frac · CIFAR-100 fine-tuning | top-1 accuracy: 78.7 -> 84.2 | Improved |
| remove random crop from pre-training augmentation (ablation) | 1p-frac · CIFAR-100 fine-tuning | top-1 accuracy: 84.2 -> 80.1 | Hurt |
| remove Mixup and Cutmix from pre-training augmentation (ablation) | 1p-frac · CIFAR-100 fine-tuning | top-1 accuracy: 84.2 -> 80.4 | Hurt |
| remove horizontal flipping from pre-training augmentation (ablation) | 1p-frac · CIFAR-100 fine-tuning | top-1 accuracy: 84.2 -> 84.3 | No clear effect |
| use affine perturbation instead of elastic perturbation for single real image (ablation) | single real image pre-training · CIFAR-100 (mean of five images) | mean accuracy: 37.0 -> 81.9 | Improved |
| use Canny edge image instead of RGB image for single real image pre-training (ablation) | single real image pre-training · CIFAR-100 (mean of five images) | mean accuracy: 81.6 -> 82.2 | Improved |
Closest papers by meaning
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.90)
- Improving the Scaling Laws of Synthetic Data with Deliberate Practice (ICML 2025 · similarity 0.87)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.87)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.87)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.86)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.86)
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.
