Research map / Image generation / Fast diffusion sampling
A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence
Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, Ming-Hsuan Yang
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 |
|---|---|---|---|
| Fuse independently L2-normalized Stable Diffusion and DINOv2 features | DINOv2-ViT-B/14 correspondence · SPair-71k test, zero-shot | PCK@0.10: 55.6 to 64.0; +8.4p | Improved |
| Fuse independently L2-normalized Stable Diffusion and DINOv2 features | DINOv2-ViT-B/14 correspondence · SPair-71k test, supervised bottleneck | PCK@0.10: 59.9 to 74.6; +14.7p | Improved |
| Fuse independently L2-normalized Stable Diffusion and DINOv2 features | DINOv2-ViT-B/14 correspondence · PF-Pascal, zero-shot | PCK@0.10: 77.3 to 86.1 | Improved |
| Fuse independently L2-normalized Stable Diffusion and DINOv2 features | DINOv2-ViT-B/14 correspondence · TSS, zero-shot | PCK@0.05 average: 72.0 to 79.7; +7.7p | Improved |
| Use Stable Diffusion features instead of DINOv2 features | Dense correspondence flow · TSS | flow smoothness (first-order difference), average: 10.15 -> 5.90 | Improved |
| Fuse Stable Diffusion and DINOv2 features | Instance swapping · 800-pair subset | Quality score: 63.18 -> 64.84 | Improved |
| Fuse Stable Diffusion and DINOv2 features | Instance swapping · 800-pair subset | CLIP score: 72.63 -> 73.21 | Improved |
| Fuse Stable Diffusion and DINOv2 features | Instance swapping · 800-pair subset | FID score (CLIP-based): 12.49 -> 12.47 | Improved |
| Use only the decoder feature rather than combining it with the encoder feature (ablation) | Stable Diffusion feature extraction · Semantic correspondence | correspondence performance | Improved |
| Use the decoder feature rather than a single convolution, self-attention, or cross-attention sub-layer feature (ablation) | Stable Diffusion feature extraction · Semantic correspondence | correspondence performance | Improved |
| Use DINOv2 token features from layer 11 instead of DINOv1 key features from earlier layers | DINO correspondence features · SPair-71k subsets, 20 pairs per category | PCK@0.10: 31.4 -> 52.7 | Improved |
Closest papers by meaning
- SD4Match: Learning to Prompt Stable Diffusion Model for Semantic Matching (CVPR 2024 · similarity 0.89)
- Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.89)
- Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations (NeurIPS 2025 · similarity 0.89)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.88)
- Text-to-Image Diffusion Models are Great Sketch-Photo Matchmakers (CVPR 2024 · similarity 0.88)
- Text-Image Alignment for Diffusion-Based Perception (CVPR 2024 · 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.
