Reproduce “Quality-Aware Language-Conditioned Local Auto-Regressive Anomaly Synthesis and Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Long Qian, Bingke Zhu, Yingying Chen, Ming Tang, Jinqiao Wang
Researchers from the Chinese Academy of Sciences developed ARAS and QARAD, a framework for synthesizing high-fidelity, language-conditioned local anomalies and integrating them into detector training using quality-aware re-weighting. This approach achieves new state-of-the-art performance on MVTec AD, VisA, and BTAD datasets, while also speeding up anomaly synthesis by 5 times compared to diffusion-based methods.
Key insights
- Token-anchored auto-regressive editing can achieve micro-structure-preserving local anomaly synthesis, critical for textured materials, by integrating context and linguistic control.
- Dynamically weighting synthetic training data based on its semantic consistency with the guiding prompt (Quality-Aware Weighting) significantly improves anomaly detector performance and stability.
- A decoupled pipeline, combining an efficient, training-free local anomaly generator with a quality-aware detector, offers both performance gains and computational efficiency without increasing inference latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.