Survival-law modeling of anomaly leakage in diffusion reconstruction
Model anomaly persistence during diffusion-based reconstruction as a deterministic recursion that yields a continuous “survival law.” The law describes how the anomaly decays as noise decreases, controlled by a retention parameter tied to which internal states the sampler visits. This framing treats leak growth with steps as a form of exposure bias and relates trade-offs between leak and reconstruction/detail error.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
It provides an explicit, step-dependent characterization of how anomalies “survive” diffusion reconstruction, enabling predictions about how changing samplers, schedules, or nearby starting conditions affects anomaly leakage and the residual map.
What changed
Instead of viewing anomaly detection purely as reconstruction error, the approach (1) derives a survival law for anomaly leak from the behavior of a first-order deterministic sampler, (2) uses that law to interpret step-wise leak as exposure bias and to study how path/scheduling/training change the leak–detail trade-off, and (3) introduces a fine-tuning strategy that performs inference-like starts and adds evidence to residuals while preserving their ranking.
Where it helps
- Explaining and predicting anomaly leakage for diffusion reconstruction pipelines
- Designing sampler/schedule choices to control anomaly “survival”
- Interpreting exposure bias in step-wise behavior of reconstruction-based detectors
- Fine-tuning reconstructor behavior using inference-like starts
- Adjusting residual-map evidence without harming residual ranking
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The source provides claims tied to specific diffusion anomaly-detection setups and reported outcomes; details needed to reproduce the method (e.g., exact architectural choices, loss formulations, or how comparator evidence is computed) are not fully specified in the provided text.
Sources
- SurvAD: How Anomalies Survive Diffusion Reconstruction ↗
arXiv · primary · Checked 2026-11-10
