EdgeZSAD: Value-Range Quantization for Deploying Zero-Shot Anomaly Detection on Edge Neural Processing Units

summary

The gist

Industrial inspection needs zero-shot anomaly detection (ZSAD) that remains useful under edge deployment constraints The resulting model reaches an average image AUROC of 91.6 on MVTec-AD and 88.2 on…

In short

EdgeZSAD investigates deploying zero-shot anomaly detection (ZSAD) models on resource-constrained edge devices like Jetson Orin Nano Super and RB5 Gen2. The system uses a compact TinyViT-21M-512 backbone combined with an asymmetric global-local readout to achieve high accuracy (AUROC 91.6/88.2) while ensuring the exported model retains its ranking behavior on the edge.

Key concepts

EdgeZSAD
A compact reference system for industrial ZSAD built around a TinyViT-21M-512 backbone and an EdgeGLR readout. It focuses on creating a deployment-constrained regime where one compact graph is trained to work across various edge runtimes.
EdgeGLR (Asymmetric Global-Local Readout)
A fusion mechanism that combines two parts: a stable global branch (GRB) predicting the overall image score, and a sparse local branch (SCB) using a mask head to predict an anomaly map. This is then reduced via sparse top-k pooling to generate a local correction term.
Real-IAD-DR Recipe
A deterministic source-side training recipe that increases training difficulty without increasing the deployed graph size. It synthesizes local defects on normal regions using a GLASS recipe and adds edge corruptions like blur and compression, shaping the model to handle deployment degradation.

Terminology used across episodes

This episode discusses

The paper

EdgeZSAD: Value-Range Quantization for Deploying Zero-Shot Anomaly Detection on Edge Neural Processing Units · Read on arXiv

Gachon University · Plaid Labs Inc.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "EdgeZSAD: Value-Range Quantization for Deploying Zero-Shot Anomaly Detection on Edge Neural Processing Units".

Jane: Here's the gist of the paper. Industrial inspection needs zero-shot anomaly detection (ZSAD) that remains useful under edge deployment constraints The resulting model reaches an average image AUROC of 91.6 on MVTec-AD and 88.2 on VisA, while remaining directly deployable on Jetson Orin Nano Super (TensorRT FP16) and RB5 Gen2 (QNN GPU FP16) Across the six device-rescored benchmarks, imageAUROC drift stays below 0.2 points, indicating that the exported graph preserves host-side ranking behavior in the evaluated deployment setting Pruning redundant reasoning tokens The study investigates EdgeZSAD, a compact reference system built around a TinyViT-21M-512 backbone, an asymmetric global-local readout (EdgeGLR), and a reproducible source-side training recipe (Real-IAD-DR) The research focuses on defining a deployment-constrained industrial ZSAD regime and instantiating it with a single compact TinyViT-21M-512 + EdgeGLR graph that remains exportable to real edge runtimes How the training signal is built The training side uses Real-IAD-DR, a deterministic source-side recipe derived from Real-IAD

10: that increases sourceside difficulty without enlarging the deployed graph This recipe reshapes source difficulty by synthesizing local defects on normal regions with a GLASS-inspired recipe and adding edge-aware corruptions such as blur, compression, illumination shift, and mild sensor noise The total training objective combines image-level supervision and pixel-level localization supervision: L = λimgLimg + λmaskLmask In the selected recipe we use λimg = 1 and λmask = 2, train with AdamW for 2000 steps, and keep an exponential moving average (EMA) of the model weights Where the gains come from The authors identify two levers that matter inside that regime: EdgeGLR shapes the image score because sparse top-k fusion preserves useful local evidence, as SCB-only on VisA falls to 79.32 AUROC versus 89.41 for GRB-only Real-IAD-DR shapes the training signal because the sourcecorpus recipe matters more once the deployed backbone is small and the model has less capacity to absorb weak supervision Furthermore, Table 3 shows that replacing Real-IAD with Real-IAD-DR improves imagelevel ranking on all four datasets and improves pixel metrics on three of four (KSDD2 pixel PRO is the exception) Accuracy–Scale–Deployment Summary EdgeZSAD occupies a much smaller parameter and FLOP region than the CLIP-L and DINOv2-L families while remaining validated on edge devices Feasibility is backend-specific rather than FLOPs-only, as the TinyViT-21M-512 + EdgeGLR path avoids failure modes like the DINOv2-L Resize operator on RB5 (Table 4) The main practical result is that compact zero-shot transfer remains viable when the source corpus, score formation, and export path are designed together from the outset Edge Device Evaluation The device table serves a different purpose from the host table by asking whether that same checkpoint survives deployment with minimal drift on the device-rescored benchmarks The Jetson and RB5 outputs remain close to the host outputs for the reported rows, and Figure 4 shows that both the ranking signal and the spatial anomaly pattern are preserved after export This distinction matters because host-side parameter or FLOP counts do not guarantee successful export Limitations Three boundary conditions remain, including localization being the main accuracy bottleneck where MVTec-AD pixel PRO (67.8) sits below the strongest baseline (88.6) Dataset behavior remains uneven, such as on DAGM where image AUROC trails the strongest reference by about 8 points Finally, deployment evidence is limited to FP16 outputs Future Work A natural extension is fixed-function NPU deployment on Hailo8

32: or DeepX DX-M1

33: , which require INT8 quantization and restricted operator sets More broadly, recent post-training quantization results for vision transformers suggest that low-bit deployment can retain useful accuracy when the quantization scheme respects transformer-specific sensitivities Other extensions include multi-seed training, learnable fusion weights, and multi-source corpus mixing to narrow the remaining synthetic-texture gap without enlarging the deployed graph We presented EdgeZSAD, a deployment-constrained ZSAD study that combines TinyViT-21M-512, EdgeGLR, and Real-IAD-DR Across three runs, it reaches an average image AUROC of 91.6 on MVTec-AD and 88.2 on VisA The main practical result is that compact zero-shot transfer remains viable when the source corpus, score formation, and export path are designed together from the outset Acknowledgments Supported by the 2026 Re:Startup Success Package (Project No.

Tom: 20396577) through the Seoul Center for Creative Economy & Innovation, Republic of Korea, and by the Gachon University research fund (GCU-202500670001) References

1: A.

Jane: Radford, J.

Tom: W.

Title and authors: Jane: The paper's title, "EdgeZSAD: Value-Range Quantization for Deploying Zero-Shot Anomaly Detection on Edge Neural Processing Units," tells us exactly what it’s trying to solve. It signals that they are focusing specifically on the deployment side of zero-shot anomaly detection.

Lu: They use value-range quantization, which is a specific way of reducing model size and computational load, which is crucial for getting it onto constrained hardware like those embedded systems. This isn't just general pruning; it’s tailored to the target chips.

Meng: So they are essentially trying to find a sweet spot where you get good zero-shot performance without needing a massive model that requires huge memory or too many complex operations for a chip with limited resources. That makes sense from my end, because deployment failures are usually about hitting those limits.

Tom: They also point out that they study this regime through EdgeZSAD, which is built around the TinyViT-21M-five hundred twelve backbone and EdgeGLR, and a reproducible source-side training recipe called Real-IAD-DR <ref:2606.16119#pg1,EdgeGLR, and a reproducible source-side training recipe>. It’s not just about the final model; it’s about the whole system designed for edge deployment.

Jane: That's right, and that means they aren't just tweaking a large model; they are designing an entire workflow—the backbone, the readout mechanism, and how you train it all together to fit that compact structure.

Lalam: The authors show that across three independent runs with this setup, the resulting model achieves an average image AUROC of ninety-one point six on MVTec-AD and eighty-eight point two on VisA, while remaining directly deployable on Jetson Orin Nano Super with TensorRT FP16 and RB5 Gen2 with QNN GPU FP16 <ref:2606.16119#pg1,an average image AUROC of 91.6 on MVTec-AD and 88>.

Tom: That's a concrete number, which is good because it shows the performance isn't just theoretical; it’s measured on the actual target hardware. But I want to know how stable that performance is when you move between different devices or benchmarks.

The paper's summary: Jane: The summary explains that they define a deployment-constrained industrial zero-shot anomaly detection regime and then instantiate it with a single compact TinyViT-21M-five hundred twelve plus EdgeGLR graph that remains exportable to real edge runtimes <ref:2606.16119#pg1>. This whole approach is key because it's about making sure the exported graph actually works on the device.

Lu: The method splits the work deliberately: at inference time, you only keep the TinyViT-21M-five hundred twelve and EdgeGLR, which is simple enough for TensorRT and QNN <ref:2606.16119#pg1>. They move all the heavy lifting to source-corpus construction and checkpoint selection during training, which changes how you train it but doesn't change the deployed operator path.

Meng: That split makes sense for me because it means we can optimize the deployment side—the inference path—to be as minimal and fast as possible, while dedicating more resources to making sure the model learns what it needs to learn from that compact structure.

Tom: They also detail how the training signal is built using Real-IAD-DR, a deterministic source-side recipe derived from Real-IAD that increases sourceside difficulty without enlarging the deployed graph. This recipe reshapes source difficulty by synthesizing local defects on normal regions with a GLASS inspired recipe and adding edge aware corruptions like blur, compression, illumination shift, and mild sensor noise.

Jane: So they aren't just training on more data; they are making the training signal reflect what the actual deployment degradation looks like in real-world scenarios. That’s a smart way to make the model robust for the edge.

Lalam: The total training objective combines image-level supervision and pixel-level localization supervision, with lambda img equals one and lambda mask equals two, they train with AdamW for two thousand steps, and keep an exponential moving average of the model weights. This shows a very specific recipe for how they tune that training process.

The paper's improvements: Tom: They identify two levers that matter inside that regime: EdgeGLR shapes the image score because sparse top-k fusion preserves useful local evidence, and SCB-only on VisA falls to seventy-nine point three two AUROC versus eighty-nine point four one for GRB-only Real-IAD-DR shapes the training signal because the source corpus recipe matters more once the deployed backbone is small and the model has less capacity to absorb weak supervision.

Lu: EdgeGLR works because it combines a stable global branch, called GRB, which applies global average pooling and a linear head to produce a stable ranking probability sg, with a sparse local branch, SCB, which uses a depthwise-separable mask head m of F4 to predict a dense anomaly map P of x. Then they reduce sigma P of x by sparse top-k pooling to get a local correction term sl.

Meng: So the fusion design is critical; it’s not just about having two heads, it’s about how you combine the stable global signal with the localized evidence, and that combination gives them stability across different datasets.

Jane: And they show that replacing Real-IAD with Real-IAD-DR improves image level ranking on all four datasets and improves pixel metrics on three of four, only leaving KSDD2 pixel PRO as an exception. That tells us that the source corpus recipe really matters for getting better localization too.

Tom: The main practical result they highlight is that compact zero-shot transfer remains viable when the source corpus, score formation, and export path are designed together from the outset. It’s a design philosophy, not just a training trick.

Conclusion: Jane: So to wrap up, they presented EdgeZSAD as this deployment-constrained zero-shot anomaly detection study that combines TinyViT-21M-five hundred twelve EdgeGLR, and Real-IAD-DR <ref:2606.16119#pg1>. The main practical result is that compact zero-shot transfer remains viable when the source corpus, score formation and export path are designed together from the outset.

Lu: Yeah, it’s a system where you design for deployment constraints from the start so you can actually get a model that runs on edge hardware without sacrificing too much accuracy. The focus on the small deployed graph versus heavy off-device training is really illuminating.

Meng: For me, what this means is that we should stop thinking about just making the largest model and then trying to shrink it down; instead, you need to design the whole system for resource limits right from day one.

Lalam: I’m glad they showed that across three runs, they hit an average image AUROC of ninety-one point six on MVTec-AD and eighty-eight point two on VisA, and that the exported graph preserves host-side ranking behavior in the evaluated deployment setting <ref:2606.16119#pg1,an average image AUROC of 91.6 on MVTec-AD and 88>.

Tom: A solid result because it proves that you can achieve good zero-shot performance even with very small backbones when you engineer the system right for edge deployment constraints. That's what EdgeZSAD is all about.

More episodes

← Home