Research map / Image generation / Fast diffusion sampling
Difficulty Controlled Diffusion Model for Synthesizing Effective Training Data
Zerun Wang, Jiafeng Mao, Xueting Wang, Toshihiko Yamasaki
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 medium-difficulty synthetic data instead of easy synthetic data | Real + synthetic training data · Imagenette | accuracy (%): 95.2 (+0.2) -> 95.8 (+0.8) | Improved |
| add extremely hard synthetic data | Real + synthetic training data · Imagenette | accuracy (%): 95.0 -> 94.6 (-0.4) | Hurt |
| use difficulty-conditioned generation instead of Real-Fake | Synthetic data generator · ImageNet (10% synthetic) | top-1 accuracy (%): 78.61 -> 78.74 | Improved |
| use difficulty-conditioned generation instead of Real-Fake | Synthetic data generator · ImageNet (100% synthetic) | top-1 accuracy (%): 78.62 -> 78.63 | No clear effect |
| fine-tune Stable Diffusion with LoRA instead of no fine-tuning | Stable Diffusion synthetic data · ImageNet (10% synthetic) | top-1 accuracy (%): 77.90 -> 78.61 | Improved |
| use difficulty-conditioned generation instead of Real-Fake with ViT-small | Synthetic data generator · Imagenette (75% synthetic) | accuracy (%): 84.8 -> 86.0 | Improved |
| use difficulty-conditioned generation instead of Real-Fake with ResNet-50 | Synthetic data generator · Imagenette (75% synthetic) | accuracy (%): 95.4 -> 96.4 | Improved |
| use difficulty-conditioned generation instead of Real-Fake with ResNet-101 | Synthetic data generator · Imagenette (75% synthetic) | accuracy (%): 95.8 -> 96.8 | Improved |
| set mean difficulty score to 0.5 rather than 0.3 (ablation) | Difficulty score sampling · Imagenette (75% synthetic) | accuracy (%): 95.8 -> 96.4 | Improved |
| raise mean difficulty score from 0.5 to 0.9 (ablation) | Difficulty score sampling · Imagenette (75% synthetic) | accuracy (%): 96.4 -> 95.2 | Hurt |
| sample difficulty with std 0.10 instead of fixed 0.00 (ablation) | Difficulty score sampling · Imagenette (75% synthetic) | accuracy (%): 95.8 -> 96.4 | Improved |
| raise difficulty std from 0.10 to 0.50 (ablation) | Difficulty score sampling · Imagenette (75% synthetic) | accuracy (%): 96.4 -> 96.0 | Hurt |
Closest papers by meaning
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.89)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.87)
- CtrLoRA: An Extensible and Efficient Framework for Controllable Image Generation (ICLR 2025 · similarity 0.86)
- Composition and Alignment of Diffusion Models using Constrained Learning (NeurIPS 2025 · similarity 0.86)
- Improving the Scaling Laws of Synthetic Data with Deliberate Practice (ICML 2025 · similarity 0.86)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 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.
