Research map / Image generation / Fast diffusion sampling
Retrieval-Augmented Diffusion Models
Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas Müller, Björn Ommer
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 |
|---|---|---|---|
| train with large diverse OpenImages database instead of small or different-domain database (ablation) | RDM on ImageNet dogs subset · ImageNet dogs | FID/precision/recall | Improved |
| use database from different domain (WikiArt) during training (ablation) | RDM on ImageNet dogs subset · ImageNet dogs | sample quality | Hurt |
| train database disjoint from train set (OpenImages) instead of train set itself, then swap inference database (ablation) | RDM · ImageNet train CLIP-FID | CLIP-FID: 8.86 -> 3.84 | Improved |
| disjoint train database (OpenImages) vs train set as database (ablation) | RDM · COCO zero-shot text-to-image | FID: 27.28 -> 22.08 | Improved |
| increase k_train from 1 to 2-8 neighbors (ablation) | RDM · ImageNet | FID / IS | No clear effect |
| increase k_train beyond 8 (to 16) (ablation) | RDM · ImageNet | recall | Hurt |
| increase k_train up to 8 (ablation) | RDM · COCO zero-shot text-to-image | FID / CLIP score | Improved |
| add classifier-free guidance (s=1.75, m=0.1) to RDM-OI (ablation) | RDM-OI · ImageNet unconditional train FID | FID: 24.50 -> 13.22 | Improved |
| top-m sampling truncation (smaller m) (ablation) | RDM-OI · ImageNet unconditional | FID: 24.50 -> 19.08 | Improved |
| replace nearest-neighbor retrieval with direct CLIP image embedding conditioning plus flow prior (ablation) | RDM inversion model · COCO zero-shot text-to-image | FID: ~87 -> ~45 | Improved |
| replace parametric LDM with retrieval-augmented RDM-OI | LDM · FFHQ | CLIP-FID: 2.12 -> 1.92 | Mixed |
| use more complex ImageNet subsets with semi-parametric models | RDM and RARM · ImageNet subsets | FID/recall | Improved |
Closest papers by meaning
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.90)
- RealRAG: Retrieval-augmented Realistic Image Generation via Self-reflective Contrastive Learning (ICML 2025 · similarity 0.89)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.89)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.89)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.88)
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.
