Research map / Image generation / GANs
DreamTeacher: Pretraining Image Backbones with Deep Generative Models
Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, Sanja Fidler
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 |
|---|---|---|---|
| use DreamTeacher feature distillation (ADM) pretraining instead of supervised ImageNet pretraining | ResNet-50 backbone · ADE20k segmentation | mIoU: 40.9 -> 42.5 | Improved |
| use ADM as teacher instead of Stable Diffusion 1.4 or eDiffI (ablation) | DreamTeacher RN50 · BDD100K segmentation | mIoU: 61.6 vs 60.2 (SD1.4) vs 58.5 (eDiffI) | Improved |
| DreamTeacher pretraining vs BYOL on in-domain nuScenes-30K | BEVFormer RN50 · nuScenes 3D detection val | NDS: 22.2 -> 28.4 | Improved |
| DreamTeacher pretraining vs BYOL on ImageNet-1k | BEVFusion RN50 camera only · nuScenes map segmentation | Mean IoU: 55.4 -> 53.9 | Hurt |
| DreamTeacher pretraining vs supervised on camera+LiDAR fusion | BEVFusion RN50 · nuScenes map segmentation C+L | Mean IoU: 59.2 -> 59.4 | No clear effect |
| increase diffusion noise step used for features (ablation) | DDPM-seg FFHQ · Sim2real MS-SynFaces-19 | mIoU: 61.5 -> 44.2 | Hurt |
| use a few diffusion steps instead of step 0 (ablation) | DDPM-seg LSUN-car · Car-20 | mIoU: 64.4 -> 75.4 | Improved |
| use DT mixed distillation with ConvNeXt-B instead of DDPM-seg | label-efficient segmentation · Bedroom-28 | mIoU: 49.4 -> 54.8 | Improved |
Closest papers by meaning
- Unlocking Pre-trained Image Backbones for Semantic Image Synthesis (CVPR 2024 · similarity 0.83)
- DreamCache: Finetuning-Free Lightweight Personalized Image Generation via Feature Caching (CVPR 2025 · similarity 0.81)
- Learning Visual Generative Priors without Text (CVPR 2025 · similarity 0.80)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.80)
- VisionDirector: Vision-Language Guided Closed-Loop Refinement for Generative Image Synthesis (CVPR 2026 · similarity 0.80)
- DreamBench++: A Human-Aligned Benchmark for Personalized Image Generation (ICLR 2025 · similarity 0.80)
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.
