Research map / Image generation / Fast diffusion sampling
Fast Diffusion-Based Counterfactuals for Shortcut Removal and Generation
Nina Weng, Paraskevas Pegios, Eike Petersen, Aasa Feragen, Siavash Arjomand Bigdeli
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 denoised-image gradients instead of nested DDPM generation in DiME | FastDiME · CelebA smile (1000 images) | total time: 12:10:08 -> 00:45:55 | Improved |
| use denoised-image gradients instead of nested DDPM generation in DiME | FastDiME · CheXpert (pacemaker) | FID: 73.2588 -> 61.4010 | Improved |
| use FastDiME instead of DiME on CelebA smile realism | DiME · CelebA smile | FID: 3.17 -> 4.18 | Hurt |
| add self-optimized masking to FastDiME (ablation) | FastDiME · CheXpert (pacemaker) | FID: 68.0411 -> 61.4010 | Improved |
| add self-optimized masking to FastDiME (ablation) | FastDiME · CheXpert (pacemaker) | L1: 0.1082 -> 0.0897 | Improved |
| replace self-optimized masking with RePaint inpainting (ablation) | FastDiME · CelebA smile subset | flip ratio: 0.998 -> 0.640 | Hurt |
| use GMD with mask instead of FastDiME gradients (ablation) | FastDiME · CelebA smile subset | flip ratio: 0.998 -> 0.881 | Hurt |
| add second pass with fixed mask (FastDiME-2+) | FastDiME · CelebA age | FID: 4.82 -> 3.60 | Improved |
| increase mask dilation from 5 to 55 (ablation) | FastDiME · CheXpert (pacemaker) | FID: 52.2663 -> 67.1333 | Hurt |
| increase mask threshold from 0.15 to 0.3 (ablation) | FastDiME · CheXpert (pacemaker) | MAD: 0.7554 -> 0.6465 | Hurt |
| train classifier on more shortcut-correlated data (D100 vs D50) | ResNet-18 shortcut detection on CheXpert · CheXpert (cardiomegaly, pacemaker) | MAD: 0.12 -> 0.36 | Improved |
| use AUROC gap between test_k and test_u to detect shortcuts | shortcut detection pipeline · CheXpert (D100) | AUROC: 0.98 -> 0.58 | Mixed |
Closest papers by meaning
- Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis (CVPR 2025 · similarity 0.85)
- Clinically-Grounded Counterfactual Reasoning for Medical Video Diagnosis (CVPR 2026 · similarity 0.85)
- MaskDiME: Adaptive Masked Diffusion for Precise and Efficient Visual Counterfactual Explanations (CVPR 2026 · similarity 0.85)
- Trade-Offs in Fine-Tuned Diffusion Models between Accuracy and Interpretability (AAAI 2024 · similarity 0.84)
- One Step Diffusion via Shortcut Models (ICLR 2025 · similarity 0.83)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.83)
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.
