NVIDIA Kumo Anomaly Brings Diffusion-Based Anomaly Detection to Multivariate Time Series

Highlights (TL;DR)

NVIDIA Kumo Anomaly (formerly known as NV-Tesseract AD-Diffusion) is a diffusion-based model for anomaly detection in multivariate time series. It combines diffusion modeling, curriculum learning, and adaptive thresholding to identify unusual behavior in noisy, high-dimensional signals.

Given a sequence of time-series observations, the library returns the original columns alongside a mean absolute error (MAE) anomaly score and a binary anomaly label for each row. Developers can choose between Segmented Confidence Sequences (SCS) and Multi-Scale Adaptive Confidence Segments (MACS) for adaptive thresholding. The model has approximately 8 million parameters and is available under the Apache 2.0 license.

Model code: NVIDIA/Kumo-TS
Model weights: nvidia/Kumo-Anomaly

The Shift to Diffusion-Based Anomaly Detection

Anomaly detection is not just a question of finding anomalous values. In a multivariate signal, an observation can become unusual because of its timing, its relationship to other measurements, or a change in the behavior of the system as a whole. Kumo Anomaly addresses precisely these challenges: noisy, high-dimensional data that changes over time and contains rare, irregular events

The model brings together three complementary components. Diffusion modeling supports time-series imputation and anomaly detection. Curriculum learning stabilizes model training. Adaptive thresholding converts anomaly scores into decisions while accounting for nonstationary behavior.

This combination separates two questions that an anomaly-detection pipeline must answer: how unusual is an observation, and when is that unusualness sufficient to raise a flag? Kumo Anomaly exposes both the continuous Mean Absolute Error (MAE) score and the resulting binary label, giving developers access to both stages of the decision.

How Kumo Anomaly Works

Kumo Anomaly’s public interface starts with a table of time-series observations and ends with scores and labels. Its main components are diffusion-based modeling, error scoring, and adaptive thresholding. It has approximately 8 million parameters with an 80/20 split per dataset.

Training data was primarily TSB-AD-M, a benchmark containing 198 curated multivariate time series from 17 public datasets. These series span multiple application domains and include both point and sequence anomalies, with variation in dimensionality, anomaly frequency, anomaly duration, and series length.

Core Concept: Diffusion-based model that learns normal time series patterns through noise corruption and reconstruction, then detects anomalies by measuring reconstruction errors. A more detailed breakdown can be found in this paper, NV-Tesseract-AD: Curriculum-Enhanced Conditional Diffusion for Multivariate Time-Series Anomaly Detection

Training Phase: Adds controlled noise to clean data to simulate missing values. The model is trained to reverse this corruption through a neural network that learns to reconstruct the original values from the noisy version.

Inference Phase: When given new, incomplete data, the model uses what it learned to recover the missing values, one step at a time, by removing the noise. In the case of anomaly detection, the model imputes all values—even observed ones—and compares the reconstructed values with the actual input. If the difference is large, that data point is flagged as a potential anomaly.

Kumo-Anomaly supports two thresholding strategies:

• SCS: Segmented Confidence Sequences.
• MACS: Multi-Scale Adaptive Confidence Segments.

To address this, NVIDIA researchers developed two patent-pending methods—Segmented Confidence Sequences (SCS) and Multi-Scale Adaptive Confidence Segments (MACS). Both are unsupervised, model-agnostic, and grounded in confidence interval theory, making them interpretable and broadly applicable beyond NV-Tesseract-AD. These methods are part of the inference process, specifically during the threshold-setting stage of the anomaly detection pipeline, where they determine when a deviation should be considered significant.

SCS divides the time series into locally stable regimes, each with its own statistical baseline. Confidence bounds adapt within each regime, ensuring sensitivity where it’s needed and restraint where natural variance is high.

MACS examines data through short-, medium-, and long-term windows simultaneously. An attention mechanism weighs the most relevant scale, while a dual-detection rule reduces the number of spurious alerts. This allows MACS to capture quick bursts and gradual drifts without requiring separate detectors.

Additional information can be found in this published paper, Segmented Confidence Sequences and Multi-Scale Adaptive Confidence Segments for Anomaly Detection in Nonstationary Time Series

Sample Use Cases:

Anomaly Detection in Semiconductor Manufacturing

In semiconductor manufacturing, the challenge is especially stark. Each wafer undergoes hundreds of precision steps, generating massive streams of sensor data, including chamber pressures, gas flows, RF power levels, and vibrations. Traditional monitoring relies on fixed thresholds and statistical process control (SPC). Alarms tend to be statistical and reactive, requiring constant recalibration as processes drift. Subtle faults or context-dependent anomalies can slip past these safeguards, and production may continue without interruption. One missed anomaly can cascade into yield losses worth millions of dollars.

This is where Kumo Anomaly proves its value. The model identifies the precise moment in a time series when an anomaly occurs, across single signals and multiple correlated sensors. This level of granularity is critical: factory facilities (fabs) can determine exactly from which point forward wafers may be faulty, take corrective action immediately, and prevent losses from cascading downstream.

By shifting from after-the-fact inspection to real-time anomaly localization, Kumo Anomaly transforms raw sensor noise into actionable insights—protecting yield, reducing costs, and laying the groundwork for more intelligent automation at fab scale. Examples of what you would see in fab data include:

• Level shifts: A sensor may suddenly shift to a higher or lower baseline, often indicating a calibration change or the start of tool drift.
• Spikes and fluctuations: Brief, erratic bursts of abnormal readings can indicate instability in plasma sources or gas flow valves. Left unchecked, these short-lived spikes can ruin an entire wafer batch.
• Sensor failures: At times, a sensor goes flat, reporting zeros or locked values that clearly diverge from reality. These “silent failures” are especially hazardous because they conceal the actual state of the process.
• Scale shifts: In other cases, sensor output suddenly jumps into an entirely different magnitude range, often due to misconfiguration or early hardware failure.

Kumo-Anomaly changes the equation, identifying the precise timestamp at which an anomaly begins. The model provides fabs with a clear decision point: wafers processed before that moment can often be saved, while those after require closer scrutiny. This granularity changes anomaly detection from a blunt “something went wrong” signal into a practical tool for managing yield.

In short, Kumo-Anomaly doesn’t replace fab alarms. It complements them with the contextual intelligence needed to act decisively at the scale of modern manufacturing.

Anomaly Detection on Databricks AI Runtime (AIR)

NVIDIA’s collaboration with Databricks gives developers a practical path to adapt Kumo Anomaly to their data using serverless NVIDIA GPUs on AI Runtime. Our research evaluation illustrates why customization matters: fine-tuning on 50,000 normal observations improved mean test average precision from 0.016 to 0.201approximately 12.5× across three paired evaluation seeds on one NVIDIA H100 GPU.

By connecting NVIDIA’s open models with governed enterprise data, MLflow experiment tracking and Unity Catalog, developers gain a repeatable foundation for exploring time-series anomaly detection. These predictive insights could complement Nemotron’s language capabilities, helping future agents investigate unusual behavior using business context. Kumo Anomaly is intended for research and development; operational benefits require use-case-specific validation.

Limitations

Data must be numeric, temporal data. There is no known limit to data set size and features. Final result depends on the thresholding method selected and users encouraged to experiment with multiple methods based on their dataset and domain.

Demo

The Kumo-TS repository provides a DataFrame-first API and automatically downloads pretrained weights from Hugging Face. The Kumo Anomaly endpoint accepts a multivariate time series as JSON or CSV. The endpoint returns a JSON or CSV array of the same length.

Start Building with Kumo Anomaly

The NVIDIA Kumo-TS repository includes inference examples, adaptive thresholding options, and a fine-tuning example using normal windows from a custom CSV

Begin with the pretrained model, inspect both its continuous scores and binary decisions, and validate behavior on data representative of the intended application.

Model code: NVIDIA/Kumo-TS
Model weights: nvidia/Kumo-Anomaly

The Kumo-TS repository also includes a fine-tuning example that uses normal windows from a user-provided CSV. This provides a documented starting point for developers who want to adapt the model to their own data rather than use only the pretrained checkpoint.