CORE: Contrast-and-Reflect for Self-Supervised Prompt Optimization in Anomaly Synthesis
Industrial anomaly detection is bottlenecked by scarce anomalous samples, which limits supervised training and defect understanding. This paper proposes CORE, a training-free framework that leverages Large Multimodal Models to synthesize realistic anomalies in both zero-shot and one-shot regimes through structured prompt optimization. The key idea is to decompose prompts into BASE, VARIANT, and NEGATIVE components and refine them via a self-supervised contrast-and-reflect mechanism without human annotations or task-specific fine-tuning. If effective, this could remove the need for costly real defect collections and enable open-world adaptation.
The paper presents an intuitively appealing idea—using LMMs as black-box generators with self-supervised prompt refinement—and reports strong downstream anomaly detection improvements on MVTec-AD and VisA. However, the evaluation leans heavily on synthetic-to-real transfer metrics while direct generation quality assessment is weak, and the contrast-and-reflect mechanism is underdeveloped experimentally with only a single optimization iteration reported and no analysis of convergence or failure modes.
The structured prompt formulation—decomposing generation into BASE, VARIANT, and NEGATIVE—is a well-motivated design that directly addresses the tension between preserving normal regions and introducing realistic defects. The authors correctly identify that general-purpose LMMs lack fine-grained defect priors, and grounding sparse user context into explicit design dimensions and constraints is a pragmatic way to bridge this gap. The downstream detection results in Table 1 are consistently strong across both zero-shot and one-shot settings, with particularly large gains in pixel-level metrics compared to prior training-free methods.
The proposed self-supervised contrast-and-reflect mechanism is the headline contribution, yet the paper provides minimal evidence that iterative optimization works beyond a single round. If one iteration suffices for cost efficiency, the closed-loop design may be unnecessary overhead, and the reader is left wondering whether additional iterations would help or degrade results.
More critically, the evaluation relies on LMM-as-a-judge for both output comparison and ground-truth alignment, but the paper never addresses known failure modes of this paradigm. Prior work on LLM-as-a-judge documents position bias, verbosity bias, and limited reasoning ability that can skew relative rankings, yet no controls for these biases are implemented. Generation quality is further assessed with Inception Score and Intra-Cluster LPIPS—metrics the authors themselves acknowledge are flawed for localized defect evaluation—while human evaluation or learned perceptual metrics for anomaly realism are absent.
The downstream AD comparison is reasonably fair: the same U-Net is trained on 100 synthetic images per defect type for all methods, and training-based few-shot approaches are excluded with a defensible overfitting argument. However, the generation-quality comparison is less convincing. On VisA zero-shot, Anomagic actually achieves a higher Inception Score than CORE (1.83 vs. 1.54), and the IC-L metric remains nearly flat across zero-shot methods, suggesting diversity gains are modest. The qualitative comparisons in Fig. 4 are helpful but cherry-picked; without error analysis or per-category breakdowns in the main text, it is unclear whether the improvements are uniform or concentrated on specific defect types.
Reproducing this work would be extremely difficult. The framework depends on future or proprietary LMM backends—Gemini-3.1-Pro, Nano Banana 2, GPT-5 Mini, and GPT-Image-2—for which no open weights or reproducible APIs are described. The paper mentions appendices for detailed scoring and refinement rules but does not include them in the main text, and no code repository is indicated. Hyperparameters such as the number of candidates $K$, temperature settings, or exact prompt templates are omitted, making independent verification impossible.
Industrial anomaly detection (AD) is bottlenecked by the scarcity of anomalous samples, making high-fidelity anomaly synthesis crucial for improving downstream performance. However, existing methods either rely on scarce real anomaly references or auxiliary data, while lacking complex reasoning and iterative optimization capabilities for generating realistic defects in open-set scenarios. To address these challenges, we propose CORE, a cost-efficient and training-free prompt optimization framework based on Large Multimodal Models (LMMs), which unifies zero-shot and one-shot anomaly synthesis. CORE first introduces an Anomaly Context Grounding module to transform sparse user context into defect-aware generation context, which is subsequently instantiated as structured anomaly prompts that disentangle normal-region preservation from anomaly-specific variation, enabling controllable and faithful defect generation. Furthermore, CORE autonomously refines prompts through a self-supervised contrast-and-reflect mechanism. This design empowers general-purpose LMMs to synthesize realistic anomalies under both zero-shot and one-shot settings without task-specific training. Extensive experiments on MVTec-AD and VisA demonstrate that CORE synthesizes realistic and diverse anomalies, yielding superior improvements in downstream AD performance. Furthermore, its model-agnostic and plug-and-play paradigm enables effortless integration with diverse LMMs, facilitating adaptation to unseen open-set scenarios.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.