Research map / Image generation / Fast diffusion sampling
Diffusion Integrated Gradients: Controllable Path Generation for Flexible Feature Attribution
Soyeon Kim, Kyowoon Lee, Jaesik Choi
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 |
|---|---|---|---|
| Replace straight-line IG path with diffusion-generated guided path | IG · Oxford-IIIT Pet (ResNet18) | DiffID: 0.3213 -> 0.5067 | Improved |
| Replace IG path with DiffIG path | IG · Mini-ImageNet (Inception V1) | DiffID: 0.3579 (AGI) vs 0.3578 (DiffIG) | No clear effect |
| DiffIG vs IG on deletion | IG · Mini-ImageNet (ResNet18) | Deletion: 0.2375 -> 0.2621 | Hurt |
| Increase sampled paths N from 1 to 50 (ablation) | DiffIG · Oxford-IIIT Pet (ResNet18) | DiffID: 0.3791 -> 0.5067 | Improved |
| Increase N from 30 to 50 (ablation) | DiffIG · Oxford-IIIT Pet (ResNet18) | DiffID: 0.4940 -> 0.5067 | No clear effect |
| Increase denoising steps from 20 to 40 (ablation) | DiffIG · Oxford-IIIT Pet (ResNet18) | DiffID: 0.5067 -> 0.5011 | No clear effect |
| Reduce denoising steps from 20 to 10 (ablation) | DiffIG · Oxford-IIIT Pet (ResNet18) | DiffID: 0.5067 -> 0.4889 | Mixed |
| Swap MAR VAE for Stable Diffusion VAE (ablation) | DiffIG · Oxford-IIIT Pet (ResNet18) | DiffID: 0.5067 -> 0.4829 | No clear effect |
| Swap MAR VAE for SD VAE (ablation) | DiffIG · Oxford-IIIT Pet (VGG16) | DiffID: 0.6368 -> 0.6772 | Improved |
| Use guided DiffIG path instead of vanilla reverse denoising path (DDPath) | DDPath · Mini-ImageNet (ResNet18) | DiffID: 0.2582 -> 0.3126 | Improved |
| Apply DiffIG to ViT-B/16 | AGI baseline · Oxford-IIIT Pet (ViT-B/16) | DiffID: 0.4033 -> 0.5711 | Improved |
| Add complexity guidance strength (ablation) | DiffIG · qualitative | sparsity of attribution | Improved |
Closest papers by meaning
- Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision (AAAI 2024 · similarity 0.88)
- Denoising Diffusion Path: Attribution Noise Reduction with An Auxiliary Diffusion Model (NeurIPS 2024 · similarity 0.88)
- From Baselines to Transport Geodesics: Axiomatic Attribution via Optimal Generative Flows (NeurIPS 2026 · similarity 0.87)
- Manifold Integrated Gradients: Riemannian Geometry for Feature Attribution (ICML 2024 · similarity 0.85)
- Path Choice Matters for Clear Attributions in Path Methods (ICLR 2024 · similarity 0.85)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.85)
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.
