Research map / Image generation / Fast diffusion sampling
Diffusion Hyperfeatures: Searching Through Time and Space for Semantic Correspondence
Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, Trevor Darrell
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 |
|---|---|---|---|
| Aggregate all layers and timesteps with learned mixing weights instead of a single layer (SD-Layer-4) | Stable Diffusion v1-5 features · SPair-71k | PCK@0.1 (img): 58.80 -> 72.56 | Improved |
| Naively concatenate all layer feature maps instead of learned weighted aggregation (ablation) | Stable Diffusion v1-5 features · SPair-71k | PCK@0.1 (img): 58.80 -> 52.12 | Hurt |
| Use many timesteps (11) instead of a single step (ablation) | Diffusion Hyperfeatures · SPair-71k | PCK@0.1 (img): 63.74 -> 72.56 | Improved |
| Pass the pruned single feature map through the learned bottleneck layer (ablation) | Layer 5 Timestep 10 feature map · SPair-71k | PCK@0.1 (img): 57.69 -> 64.02 | Improved |
| Soft-mix many feature maps instead of pruning to the single best one (ablation) | Diffusion Hyperfeatures · SPair-71k | PCK@0.1 (img): 64.02 -> 72.56 | Improved |
| Swap Stable Diffusion v1-5 for v2-1 as feature source (ablation) | Diffusion Hyperfeatures · SPair-71k | PCK@0.1 (img): 72.56 -> 70.74 | Mixed |
| Increase diffusion timesteps used from 1 to 10 (ablation) | Diffusion Hyperfeatures · SPair-71k | PCK@0.1 (img): 64.61 -> 72.00 | Improved |
| Train aggregation network on DINOv2 features instead of raw layer | DINOv2 · SPair-71k | PCK@0.1 (img): 60.14 -> 68.37 | Improved |
| Concatenate all DINO layers (ablation) | DINO · SPair-71k | PCK@0.1 (img): 51.68 -> 20.17 | Hurt |
| Train aggregation network on DINO features | DINO · SPair-71k | PCK@0.1 (img): 51.68 -> 54.69 | Improved |
| Use inversion instead of generation features for real images (ablation) | Layer 4 raw features · SPair-71k late timesteps | PCK@0.1 (img): up to 5% PCK | Improved |
Later papers that compared with this method
1 later accepted paper reports Diffusion Hyperfeatures as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- MMDiff: Extending Diffusion Transformers for Multi-Modal Generation (ECCV 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations (NeurIPS 2025 · similarity 0.88)
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.87)
- Diffusion 3D Features (Diff3F): Decorating Untextured Shapes with Distilled Semantic Features (CVPR 2024 · similarity 0.87)
- EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models (ICLR 2024 · similarity 0.87)
- CleanDIFT: Diffusion Features without Noise (CVPR 2025 · similarity 0.86)
- Diffusing Differentiable Representations (NeurIPS 2024 · 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.
