🤖 machine learning

GuidedFlow: An Attention-Guided Framework for Anomaly Detection in Additive Manufacturing

This paper proposes GuidedFlow, a novel attention-guided normalizing flow framework that leverages a pre-trained ResNet and a Spatio-Temporal Attention Network to effectively detect and localize tiny defects in Additive Manufacturing, outperforming state-of-the-art models on both a new AM3D-AD dataset and the MVTec-AD benchmark.

Original authors: Sosmita Paul, Krishna Roy

Published 2026-08-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sosmita Paul, Krishna Roy

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern industrial world, the way we build things is undergoing a quiet revolution. Instead of cutting material away from a solid block, machines now build objects layer by layer, adding material only where it is needed. This process, known as additive manufacturing, allows for incredible complexity and customization, creating everything from medical implants to aerospace components. However, this precision comes with a significant challenge: ensuring that every single layer is perfect. Even a tiny flaw, such as a microscopic string of plastic or a subtle misalignment in the layers, can weaken the final object, leading to failure or safety risks. Because these defects can appear in unpredictable ways and often look very similar to the normal texture of the material, spotting them by eye is difficult and slow. To solve this, engineers have turned to computers, teaching them to watch the printing process and identify anything that looks wrong. The goal is to create a system that can see the invisible, distinguishing a perfect print from a flawed one with the same reliability a human expert would, but without getting tired or distracted.

Researchers at the New Mexico Institute of Mining and Technology have developed a new computer system designed to meet this exact challenge. They call their creation GuidedFlow. While previous computer systems tried to learn what a "normal" object looks like by memorizing patterns or trying to rebuild images from scratch, they often struggled with the tiny, messy defects common in 3D printing. These older systems sometimes missed the small problems or got confused by the complex textures of the materials. The new approach takes a different path. Instead of just looking at a single image in isolation, GuidedFlow watches the printing process as a moving story. It pays close attention to both the shape of the object and how that shape changes over time as the printer head moves. By focusing on the most important parts of the image and understanding how they relate to one another across a sequence of frames, the system becomes much better at spotting the subtle signs of a defect.

The researchers tested this system on a new collection of data they created specifically for this work, which they named AM3D-AD. This dataset includes hundreds of images and videos of real 3D-printed objects, such as bolts, gears, and cubes, made from a strong nylon material. To create the "bad" examples for the test, the team intentionally sabotaged the computer instructions that control the printer, introducing random movements that caused surface distortions and jagged paths. They also tested the system on a well-known public dataset of industrial images to see if it could handle different types of objects. The results were promising. When asked to find the flawed items, GuidedFlow correctly identified the anomalies more often than other leading computer models. It achieved a high score in distinguishing good prints from bad ones, and it was particularly good at pinpointing exactly where the defect was located on the object.

What makes this system work so well is a two-part strategy that mimics how a human might inspect a complex object. First, the system uses a pre-trained brain, a type of neural network called ResNet, to understand the basic shapes and textures in the images. Then, it adds a special attention mechanism. Imagine a security guard watching a video feed; they do not stare at every pixel with equal intensity. Instead, they focus their eyes on the areas that seem unusual or important. GuidedFlow does something similar. It learns to ignore the boring, repetitive parts of the image and concentrates its energy on the areas where something might be wrong. This "attention" is applied in two ways: it looks at the spatial details within a single frame to find tiny flaws, and it looks at the temporal details across multiple frames to see if the object is behaving consistently as it is being printed.

The researchers found that combining these two types of focus was crucial. When they tested the system without the ability to look at the movement over time, it was still good, but not as accurate. When they removed the ability to focus on specific spatial details, the system struggled to find the tiny defects. But when both were working together, the system reached its peak performance. In tests involving video clips of the printing process, the system detected anomalies with a high degree of accuracy, often outperforming other advanced methods that rely on different mathematical approaches. The system was particularly effective at finding defects in objects like gears and blocks, where the flaws were subtle and hard to see.

However, the researchers are careful to note that the system is not perfect. There were a few specific types of objects, such as certain cubes and toothbrushes, where the system was slightly less effective. In these cases, the defects were so small or the textures so repetitive that the computer had a harder time telling the difference between a flaw and a normal part of the design. This suggests that while the system is a major step forward, there is still room for improvement, especially when dealing with materials that have very complex or uniform patterns. The team also acknowledges that their testing was limited to the datasets they had available, and future work will need to expand to include a wider variety of materials and printing conditions to ensure the system is robust enough for all industrial settings.

The development of GuidedFlow represents a significant move toward safer and more reliable manufacturing. By teaching computers to watch the printing process with a focused, intelligent eye, the researchers have provided a tool that can catch errors that might otherwise slip through. This technology does not just replace human inspection; it enhances it, offering a way to monitor high-speed production lines with a level of detail that is difficult for humans to maintain over long periods. As additive manufacturing continues to grow, tools like this will be essential for ensuring that the complex parts of our future are built with the highest possible quality. The work demonstrates that by combining the power of deep learning with a smart way of focusing attention, we can solve some of the most stubborn problems in modern engineering.

Technical Summary: GuidedFlow for Anomaly Detection in Additive Manufacturing

Problem Statement
Additive Manufacturing (AM) faces significant challenges in quality control due to printing defects (e.g., stringing, tiny layer inconsistencies) and potential cyber-physical intrusions. While image and video-based anomaly detection is a primary solution, existing methods exhibit limitations. Reconstruction-based approaches (e.g., autoencoders, GANs) often miss subtle or localized defects due to bias toward dominant regular patterns. Embedding-based methods struggle with context-dependent anomalies. Although normalizing flow-based methods offer exact likelihood estimation and better generalization for unforeseen defects, they frequently fail to detect tiny or irregular defects common in AM, particularly in small-data settings. Furthermore, most existing approaches treat frames independently, neglecting the temporal dynamics critical for accurate localization and early detection in dynamic 3D printing processes.

Methodology: GuidedFlow
The authors propose GuidedFlow, a novel attention-guided normalizing flow framework designed for anomaly detection and localization in both images and videos. The architecture integrates a pre-trained ResNet-50 backbone with a Spatio-Temporal Attention Network (SAN) and a Guided Spatio-Temporal Flow (GSTF) module.

Feature Extraction: A ResNet-50 backbone extracts hierarchical features ( $f_1$ to $f_4$ ) from input video frames. During training, the lower layers ( $f_1, f_2$ ) are frozen to preserve general low-level representations, while higher layers ( $f_3, f_4$ ) are fine-tuned. These features, combined with the original frame and a scene embedding, form the input for the attention network.
Spatio-Temporal Attention Network (SAN): To prioritize relevant contextual cues, the model employs a multi-head self-attention mechanism. This module transforms input features into query, key, and value vectors, generating a unified attention output vector ( $a$ ) that captures spatial and temporal dependencies across frame sequences. This conditioning vector guides the subsequent flow processing.
Guided Spatio-Temporal Flow (GSTF): The core of the framework is a hierarchical, invertible transformation that models the distribution of multi-scale feature pyramids rather than raw pixels.
- Spatial Flow: Utilizes a Laplacian pyramid decomposition to process features across multiple scales ( $L$ levels). It applies invertible coupling operations guided by the attention vector $a$ to transform pyramid levels.
- Temporal Flow: Extends the spatial transformations across the temporal dimension ( $T$ frames), modeling dependencies between sequential frames using 3D coupling operations also conditioned on $a$ .
- Mechanism: The flow uses affine coupling layers where the scale and shift parameters are modulated by a hypernetwork that maps the SAN context to the coupling parameters. This ensures the transformation is invertible, allowing for exact likelihood evaluation.
Anomaly Scoring and Localization: Anomalies are detected by evaluating the negative log-likelihood of the latent variables given the attention conditioning. Localization maps are generated by aggregating per-location contributions from the latent energy and log-determinant terms across pyramid levels and upsampling to the input resolution.

Key Contributions
The paper outlines four primary contributions:

GuidedFlow Framework: The development of a conditioned-flow framework integrating attention-guided spatial and temporal flow for image and video anomaly detection.
Spatio-Temporal Attention Network (SAN): The introduction of a multi-head self-attention module designed to enhance sensitivity to spatial and contextual anomalies.
AM3D-AD Dataset: The proposal of a new real-world dataset containing benign and anomalous images and videos of 3D-printed objects (Bolt, Gear, Nut, Block, Cube) printed with NylonX Carbon Fiber. Anomalies were generated via intentional G-code sabotage to simulate defects like surface distortions.
Comprehensive Evaluation: A rigorous evaluation of GuidedFlow on both the proprietary AM3D-AD dataset and the standard MVTec-AD industrial anomaly detection dataset.

Experimental Results
The authors evaluated GuidedFlow against state-of-the-art (SOTA) models, including DiffusionAD, INP-Former, and PyramidFlow.

Image Detection: On the AM3D-AD and MVTec-AD datasets, GuidedFlow achieved an average Image-level AUROC (I-AUROC) of 97.4% and Pixel-level AUROC (P-AUROC) of 96.8%, outperforming the compared SOTA models. It achieved peak scores of 98.7% (AM3D Bolt) and 99.3% (MVTec MetalNut).
Video Detection: When applied to video data, the inclusion of temporal attention and flow significantly improved performance. For example, the Gear, Cube, and Block categories achieved V-AUROC scores of 98.9%, 97.2%, and 98.6%, respectively.
Ablation Studies: Experiments confirmed that the combination of spatial and temporal attention with spatio-temporal flow yields the highest performance (V-AUROC 98.7%). The study highlighted that spatial flow contributes more to fine-grained localization, while temporal flow improves overall detection by incorporating inter-frame context.
Limitations: The authors note that performance decreases slightly on classes with minimal pixel-wise deviations or highly repetitive textures (e.g., Cube, Block, Toothbrush, Tile), where the model may emphasize global texture over subtle structural defects.

Significance and Claims
The paper claims that GuidedFlow addresses the specific limitations of existing flow-based methods in Additive Manufacturing by effectively handling context-dependent anomalies and small-data settings. By leveraging spatio-temporal attention to guide the normalizing flow, the model enhances sensitivity to tiny and localized defects (such as stringing) that are often missed by other approaches. The authors assert that the framework demonstrates robust generalization across diverse object geometries and defect patterns, offering a viable solution for reliable quality assurance and safety in AM deployment. The work is supported by the National Science Foundation (Award No. 2417062).

Technical Summary: GuidedFlow for Anomaly Detection in Additive Manufacturing

More like this