Research map / Image generation / Fast diffusion sampling
Revelio: Interpreting and leveraging semantic information in diffusion models
Dahye Kim, Xavier Thomas, Deepti Ghadiyaram
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 up_ft1 decoder features instead of bottleneck features (ablation) | Diff-C · Oxford-IIIT Pet (SD 1.5, t=25) | top-1 accuracy: 69.97 -> 88.61 | Improved |
| use up_ft1 for coarse-grained datasets instead of bottleneck (label purity) (ablation) | k-SAE · Caltech-101 | sigma_label: 9.35 -> 21.33 | Hurt |
| extract features at t=25 rather than t=200 (ablation) | Diff-C up_ft1 · Oxford-IIIT Pet (SD 1.5) | top-1 accuracy: 86.89 -> 88.61 | Improved |
| use t=200 features rather than t=25 for coarse dataset (ablation) | k-SAE bottleneck · Caltech-101 | sigma_label: 9.35 -> 8.17 | Improved |
| use SD 1.5 instead of SD 2.1 (ablation) | Diff-C up_ft1 · Oxford-IIIT Pet (t=25) | top-1 accuracy: 84.74 -> 88.61 | Improved |
| add CLIP-inferred class prompt as text conditioning (ablation) | Diff-C · Oxford-IIIT Pet | top-1 accuracy: 88.69 -> 90.97 | Improved |
| add CLIP-inferred class prompt as text conditioning (ablation) | Diff-C · FGVC-Aircraft | top-1 accuracy: 65.07 -> 64.98 | No clear effect |
| use a generic prompt 'A photo of a pet' instead of empty prompt (ablation) | Diff-C up_ft1 · Oxford-IIIT Pet | top-1 accuracy: -3.14 | Hurt |
| interleave diffusion up_ft1 features (t=25) with CLIP features | LLaVA-Lightning · LLaVA-Bench (In-the-Wild) | relative score: 56.6 -> 59.9 | Improved |
| interleave DINO-v2 features with CLIP | LLaVA-Lightning · LLaVA-Bench (In-the-Wild) | relative score: 56.6 -> 47.0 | Hurt |
| add diffusion features (t=25) to CLIP in LLaVA | LLaVA-Lightning · MM-Vet | correctness score: +1.1 | Improved |
| use features from pixel-space DeepFloyd-IF at up_ft0 instead of SD 1.5 | Diff-C · Oxford-IIIT Pet and Caltech-101 (t=25) | top-1 accuracy: +1.74% Pet, +2.85% Caltech-101 | Improved |
Closest papers by meaning
- CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models (NeurIPS 2026 · similarity 0.90)
- CleanDIFT: Diffusion Features without Noise (CVPR 2025 · similarity 0.89)
- Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · similarity 0.89)
- DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing (NeurIPS 2026 · similarity 0.88)
- Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.88)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 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.
