Research map / Label-efficient and robust learning / Semi-supervised learning
Diffusion Models and Semi-Supervised Learners Benefit Mutually with Few Labels
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| train conditional diffusion model on classifier pseudo-labels (1 label/class) vs unconditional | U-ViT-Large · ImageNet 256x256 generation | FID: 27.99 -> 4.34 | Helped |
| increase labels from 1 to 2 per class for pseudo-label source | DPT (U-ViT-Huge) · ImageNet 256x256 generation | FID: 3.08 -> 2.52 | Helped |
| add pseudo images from diffusion model to classifier training | MSN ViT-B/4, 2 labels/class · ImageNet val | top-1 accuracy: 64.9 -> 69.5 | Helped |
| add pseudo images from diffusion model to classifier training | MSN ViT-B/4, 1 label/class · ImageNet val | top-1 accuracy: 52.9 -> 58.6 | Helped |
| add pseudo images from diffusion model to classifier training | MSN ViT-L/7, 5 labels/class · ImageNet val | top-1 accuracy: 72.0 -> 73.4 | Helped |
| add pseudo images from diffusion model to classifier training | Semi-ViT ViT-Huge, 1% labels · ImageNet val | top-1 accuracy: 79.4 -> 80.2 | Helped |
| add pseudo images from diffusion model to classifier training | FreeMatch, 4 labels/class · CIFAR-10 | error rate: 4.93 -> 4.68 | Helped |
| retrain generator on pseudo-labels refined by DPT-augmented classifier (ablation) | DPT · ImageNet 256x256 generation, 1 label/class | FID: 4.34 -> 4.00 | Helped |
| use more pseudo images per class K (beyond 128, up to 1280) (ablation) | DPT (MSN ViT-L/7) · ImageNet val 2 and 5 labels/class | top-1 accuracy | Hurt |
| add pseudo images with K in {12,128,256,512,1280} vs none (K=0) (ablation) | DPT (MSN ViT-L/7) · ImageNet val | top-1 accuracy | Helped |
| tune classifier-free guidance scale (0.4 at 256, 0.8 at 128, 0.7 at 512) (ablation) | DPT pseudo image generation · ImageNet FID and accuracy | FID-50K / top-1 accuracy | Helped |
| use ImageNet 128x128 DPT with 1 label/class vs S3GAN with 20% labels | semi-supervised generation · ImageNet 128x128 | FID-50K: 7.7 -> 4.59 | Helped |
Closest papers by meaning
- Barely-Supervised Learning: Semi-supervised Learning with Very Few Labeled Images (AAAI 2022 · similarity 0.89)
- Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning (NeurIPS 2024 · similarity 0.88)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.87)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.87)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.86)
- Non-Stationary Predictions May Be More Informative: Exploring Pseudo-Labels with a Two-Phase Pattern of Training Dynamics (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.