Research map / Image generation / Fast diffusion sampling
Denoising Diffusion Path: Attribution Noise Reduction with An Auxiliary Diffusion Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Denoising Diffusion Path (DDPath), which builds an attribution path from samples of a classifier-guided diffusion model centered on the target image, and integrates it with IG, BlurIG and GIG to reduce attribution noise.
Problem
Path-based attribution methods like integrated gradients accumulate noise along paths whose intermediate images drift away from the training data distribution, distorting explanations.
Approach
Replace hand-designed interpolation paths with a piece-wise linear path formed by reverse diffusion samples from a classifier-guided diffusion model, with sampling mean centered on the target image and scaled classifier guidance (rho = 1 - t/T, kappa = t/T), so intermediate images stay near the data distribution and noise decreases progressively.
Key results
- On ImageNet with ResNet-50, DDPath-IG reaches Insertion 45.1 and Overall 32.4 versus 32.5 Insertion and 18.3 Overall for IG.
- Pointing game on COCO: DDPath-BlurIG reaches 50.0% on ResNet-50 versus 46.6% for BlurIG.
- Ablation shows the default scaling (rho = 1 - t/T) outperforms the reversed scaling, e.g., DDPath-IG AIC 18.9 vs 16.2 on VGG-19.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the linear IG path with a denoising diffusion path | IG · ImageNet-1k validation, VGG-16 | Insertion: 21.7 -> 25.9 | Helped |
| replace the linear IG path with a denoising diffusion path | IG · ImageNet-1k validation, VGG-19 | Insertion: 23.2 -> 27.8 | Helped |
| replace the linear IG path with a denoising diffusion path | IG · ImageNet-1k validation, ResNet-50 | Insertion: 33.8 -> 45.1 | Helped |
| replace the linear IG path with a denoising diffusion path | IG · ImageNet-1k validation, VGG-19 | Deletion: 13.5 -> 12.1 | Helped |
| replace the linear IG path with a denoising diffusion path | IG · ImageNet-1k validation, ResNet-50 | Deletion: 13.5 -> 12.7 | Hurt |
| use DDPath counterparts instead of baseline path methods | IG, BlurIG, and GIG · MS COCO 2017 validation | Pointing game: IG: 42.3 -> 44.7, 43.4 -> 45.2, 45.2 -> 46.9; BlurIG: 45.0 -> 47.2, 45.3 -> 48.9, 46.6 -> 50.0; GIG: 45.2 -> 46.3, 44.9 | Helped |
| reverse the default scaling scheme (ablation) | DDPath-IG · ImageNet validation | AIC: 18.9 -> 16.2 | Hurt |
| reverse the default scaling scheme (ablation) | DDPath-IG · ImageNet validation | Deletion: 12.1 -> 14.5 | Hurt |
| use scaling exponent a = 2.0 rather than a = 1.0 (ablation) | DDPath-IG · ImageNet validation, VGG-19 target model | Insertion: 45.3 -> 48.9 | Helped |
| add manually sampled noise to the baseline attribution path (ablation) | IG, BlurIG, and GIG · ImageNet validation | Insertion: IG: 21.7 -> 23.5; BlurIG: 19.2 -> 20.6; GIG: 21.2 -> 21.0 | Mixed |
| add manually sampled noise to the baseline attribution path (ablation) | IG, BlurIG, and GIG · ImageNet validation | Deletion: IG: 14.8 -> 13.3; BlurIG: 13.3 -> 13.6; GIG: 14.1 -> 12.5 | Mixed |
| increase DDPath sampling steps (ablation) | DDPath-IG · ImageNet-1k validation | Insertion | Helped |
Surprising findings
- DDPath performs worse than baselines at early path steps and only gains with longer paths, with the baselines gaining less from extra steps.
- DDPath-IG slightly outperforms IG-Noise variants, suggesting that simply adding noise along the path is not sufficient for noise reduction.
Limitations
- Requires many sampling steps (250), making it slower than existing path methods.
- Performs worse than baselines at early steps and deletion scores are not always best.
- Code is not yet released; evaluation limited to image classifiers (VGG and ResNet) on ImageNet and COCO.
Experimental setup
Datasets: ImageNet-1k MS COCO 2017
Benchmarks: ImageNet-1k validation (insertion/deletion, AIC) MS COCO 2017 validation (pointing game)
Metrics: insertion AUC deletion AUC overall (insertion minus deletion) AIC pointing game hit rate
Baselines: IG SmoothGrad BlurIG GIG Guided-BP Smooth IG
Builds on: integrated gradients BlurIG guided integrated gradients classifier-guided diffusion SmoothGrad
Closest papers by meaning
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.89)
- Optimizing for the Shortest Path in Denoising Diffusion Model (CVPR 2025 · similarity 0.89)
- Interpreting and Improving Diffusion Models from an Optimization Perspective (ICML 2024 · similarity 0.89)
- Denoising Diffusion Step-aware Models (ICLR 2024 · similarity 0.89)
- Directly Denoising Diffusion Models (ICML 2024 · similarity 0.88)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (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.