DeltaDeno: Zero-Shot Anomaly Generation via Delta-Denoising Attribution
This addresses the challenge of anomaly generation for computer vision applications where anomaly data is scarce, offering a training-free solution.
The paper tackled the problem of generating anomalies without real samples or training by proposing DeltaDeno, a zero-shot method that uses diffusion denoising to localize and edit defects, resulting in improved downstream detection performance on public datasets.
Anomaly generation is often framed as few-shot fine-tuning with anomalous samples, which contradicts the scarcity that motivates generation and tends to overfit category priors. We tackle the setting where no real anomaly samples or training are available. We propose Delta-Denoising (DeltaDeno), a training-free zero-shot anomaly generation method that localizes and edits defects by contrasting two diffusion branches driven by a minimal prompt pair under a shared schedule. By accumulating per-step denoising deltas into an image-specific localization map, we obtain a mask to guide the latent inpainting during later diffusion steps and preserve the surrounding context while generating realistic local defects. To improve stability and control, DeltaDeno performs token-level prompt refinement that aligns shared content and strengthens anomaly tokens, and applies a spatial attention bias restricted to anomaly tokens in the predicted region. Experiments on public datasets show that DeltaDeno achieves great generation, realism and consistent gains in downstream detection performance. Code will be made publicly available.