Research map / Architectures and efficiency / Pruning and efficient tuning
Pre-training Vision Transformers with Very Limited Synthesized Images
Ryo Nakamura, Hirokatsu Kataoka, Sora Takashima, Edgar Josafat Martinez Noriega, Rio Yokota, Nakamasa Inoue
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 |
|---|---|---|---|
| replace 1,000 instances per category with a single instance plus data augmentation (OFDB vs FractalDB) | FDSL pre-training of ViT-T · CIFAR-100 | accuracy: 81.6 -> 84.0 | Improved |
| add random pattern augmentation (ablation) | 2D-OFDB-1k · CIFAR-100 | accuracy: 84.0 -> 85.3 | Improved |
| add IFS parameter fluctuation augmentation (ablation) | 2D-OFDB-1k · CIFAR-100 | accuracy: 84.0 -> 81.6 | Hurt |
| add rotation augmentation (ablation) | 2D-OFDB-1k · CIFAR-100 | accuracy: 84.0 -> 84.1 | No clear effect |
| scale 2D-OFDB to 21k categories | ViT-B · ImageNet-1k fine-tuning | accuracy: 0.4 points higher than ImageNet-21k | Improved |
| scale 2D-OFDB to 21k categories | ViT-T · ImageNet-1k fine-tuning | accuracy: 73.8 vs 74.1 (ImageNet-21k) | Hurt |
| reduce instances per category from 1000 to 1 (ablation) | ImageNet-1k pre-training (supervised) · CIFAR-100 | accuracy | Hurt |
| use fixed 30-degree yaw-only viewpoints for 3D fractals instead of random multi-axis viewpoints (ablation) | 3D-OFDB · CIFAR-100 | accuracy | Improved |
| balanced easy:hard (50:50) data pruning for category selection (ablation) | 2D-OFDB-1k from 21k · CIFAR-100 | accuracy | Improved |
| convert one-instance ImageNet to binary or Canny images (ablation) | ImageNet-1k subset pre-training · CIFAR-100 | accuracy | Improved |
Closest papers by meaning
- Replacing Labeled Real-Image Datasets With Auto-Generated Contours (CVPR 2022 · similarity 0.92)
- Pre-training Vision Models with Mandelbulb Variations (CVPR 2024 · similarity 0.89)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.84)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.84)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.83)
- Should VLMs be Pre-trained with Image Data? (ICLR 2025 · similarity 0.83)
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.
