Research map / Vision-language and multimodal / CLIP and zero-shot

Align3D-AD: Cross-Modal Feature Alignment and Dual-Prompt Learning for Zero-shot 3D Anomaly Detection

NeurIPS 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Attribute-based zero-shot learning · 0 citations (OpenAlex)

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 Align3D-AD, which uses RGB images from auxiliary categories during training to align rendering features of point clouds into the RGB CLIP feature space, then learns modality-specific prompts with a dual-prompt contrastive loss for zero-shot 3D anomaly detection.

Problem

Zero-shot 3D anomaly detection relies on multi-view renderings of point clouds fed to RGB-pretrained vision-language encoders (CLIP), creating a domain gap because renderings mainly encode geometry rather than appearance semantics.

Approach

Use RGB images of auxiliary categories only during training to learn a feature aligner that maps rendering features into the RGB semantic space, reweight local alignment by patch-level cross-modal consistency, and learn separate text prompts for RGB-aligned and rendering branches with a contrastive alignment between them; inference uses rendering only.

Key results

• On MVTec3D-AD one-vs-rest, Align3D-AD reaches 83.0 O-R, 94.3 O-A, 95.9 P-R and 86.5 P-P, versus 82.0 O-R for PointAD.
• On Eyecandies one-vs-rest, it reaches 71.3 O-R and 76.6 O-A, above PointAD's 69.1 and 73.8.
• In the cross-dataset setting trained on MVTec3D-AD, it reaches 76.1 O-R on Real3D-AD versus 75.9 for PointAD.

Surprising findings

• RGB-aligned features alone (80.1 O-R on MVTec3D-AD) are slightly worse than rendering-only features (82.5), yet combining them gives the best results.

Limitations

• Added RGB branch and modality-specific prompt learners increase model complexity.
• Performance depends on rendering quality; missing or noisy renderings degrade alignment.
• Hyperparameters may need per-category fine-tuning.

Experimental setup

Datasets: MVTec3D-AD Eyecandies Real3D-AD

Benchmarks: MVTec3D-AD Eyecandies Real3D-AD

Metrics: AUROC AP PRO

Baselines: CLIP AA-CLIP MVP-PCLIP GS-CLIP PointAD ZUMA

Builds on: PointAD CLIP AnomalyCLIP WinCLIP

Closest papers by meaning

• PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection (NeurIPS 2024 · similarity 0.90)
• Looking 3D: Anomaly Detection with 2D-3D Alignment (CVPR 2024 · similarity 0.87)
• Beyond Single-Modal Boundary: Cross-Modal Anomaly Detection through Visual Prototype and Harmonization (CVPR 2025 · similarity 0.87)
• Escaping Plato's Cave: Towards the Alignment of 3D and Text Latent Spaces (CVPR 2025 · similarity 0.87)
• Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.86)
• Uni$^2$Det: Unified and Universal Framework for Prompt-Guided Multi-dataset 3D Detection (ICLR 2025 · similarity 0.86)