What’s New In Wowza Video Intelligence Framework (VIF)? (Updated)

See what’s new in Wowza’s Video Intelligence Framework (VIF) and how the latest version helps teams run intelligent live and on-demand video workflows on hardware they own today, without compromising reliability.

What’s New In VIF 1.2?

Wowza Video Intelligence Framework (VIF)® 1.2 builds on the existing framework, adding model chaining and motion detection as new capabilities for all users. Model chaining uses the result of one analysis stage to determine the next steps, so a detection in one stage can trigger a second or third model, such as object detection or a vision-language model (VLM). Motion detection tracks and identifies whether motion is occuring in a video feed, and because it can run on just the CPU, it does not require a GPU. Together, the two features let teams reserve GPU-intensive analysis for the moments when something actually happens in the frames. VIF 1.2 runs on Wowza Streaming Engine 4.11 or later. Full details are in the VIF release notes.

Wowza’s Video Intelligence Framework (VIF) is the intelligence layer for Wowza Streaming Engine that turns live and recorded video into structured, actionable detection data on infrastructure the organization already controls. Customers with active maintenance and support plans benefit from regular updates to functionality and stability improvements, including these new updates in VIF version 1.2. This page covers what is in the latest release, as well as what earlier updates such as VIF version 1.1 introduced.

Since launch, VIF was named one of three finalists for Streaming Media’s Streaming Innovation of the Year for the 2026 Readers’ Choice Awards. VIF, integrated with the NVIDIA Synthetic Video Detector (SVD), was also shortlisted for the IAMT Impact Awards 2026 in the Products, Services, and Monetization category.

How Does Model Chaining Work In VIF?

Model chaining in Wowza Video Intelligence Framework (VIF) links analysis stages together so detections from one model can trigger further analysis by a second or third model. Lightweight detectors, such as motion detection, can gate more compute-intensive models so they run only when needed. This helps optimize and reduce unnecessary processing and inference costs. A chain that starts with motion detection, for example, can run on an edge device using just the CPU, and then only pass frames to object detection when motion appears. Then, the models could pass those frames to a VLM, such as NVIDIA Nemotron or NVIDIA Cosmos, to generate a scene description if a certain type or class of object is detected.

What Models Are Supported In VIF Model Chains?

VIF can chain any supported model with any other supported model. That includes the new motion detection model, the default RF-DETR object detection reference model, a custom-trained object detection model, or a vision-language model, including Qwen3-VL, NVIDIA Nemotron, and NVIDIA Cosmos reasoning models. The NVIDIA Synthetic Video Detector (SVD) is also available as a chaining stage for on-demand (VOD) analysis only.

Teams can build their own custom model chains directly in the VIF Manager user interface. Here, developers can compose stages, select an execution mode, and set sensitivity or confidence thresholds. Everything available in the Manager UI is also configurable through the REST API. For guidance on choosing a model for each stage, read our blog that covers how to configure custom AI analysis in VIF.

Model chaining extends VIF’s model flexibility and lets teams arrange analysis stages in the order their workflow requires. There are three ways a model chain can be executed:

1. Linear mode runs every stage in order, on each analyzed video window.
2. Conditional mode uses developer-defined rules to decide which stage runs next.
3. Dynamic mode passes decisions to a custom event listener, written by the developer or client, to pick which stage runs next.

What Are Some Example Model Chains For Wowza VIF?

There are a number of potential model chains that can be tailored to certain industry applications and use cases, from transportation monitoring, to perimeter security, to content preparation, and everything in between.

In a transportation monitoring use case, a transit platform after hours would employ motion detection to watch feeds on CPU-powered cameras. When motion appears, VIF (deployed on the camera itself) sends the frames to an object detection model to identify if someone left a bag unattended, or if someone has sustained an injury, or if some criminal activity might be taking place. VIF sends the frame to a VLM to describe the scene, operators receive webhooks with the description, and they decide what to do next. This all takes place in seconds, not minutes or hours, and the GPU only wakes up when something is happening.

Expensive models run only when an earlier stage calls for them, so the final result is richer and more actionable than any single model produces on its own.

Other examples of model chains in VIF include:

• Perimeter Security: Use motion detection to trigger object detection when monitoring a prohibited area. Feed that to a VLM for a detailed description. The chain runs on CPU-only edge cameras, only wakes on movement, then routes to a GPU.
• Facility Monitoring: For hazardous worksites, use object detection to trigger a conditional workflow. If a person is detected near a restricted zone, it routes to a VLM for a behavior description and triggers a webhook. If a forklift is detected, log the output without invoking the second-stage VLM.
• Live Sports/Broadcasting: Clipping and monetization get much easier. Scene analysis triggers object detection, which triggers a VLM description. Identify an in-game moment, confirm which players are involved, and generate a natural-language description for metadata, captions, or tagging.
• Content Review/Audits: For content authenticity and AI-generated media disclosures, Verify is a model chain for VIF we built with NVIDIA. Feed VOD libraries into the NVIDIA Synthetic Video Detector (SVD). If a synthetic clip is flagged, feed the output into a VLM for further analysis and validation before any video gets published.

How Does VIF Detect Motion?

VIF 1.2 adds video motion detection (VMD) for live camera feeds. Motion detection can act as the first stage of a chain, or it can run on its own, CPU-only. As the first stage in a model chain, motion detection holds further analysis until motion appears in the frame. As a result, static scenes consume no inference capacity. That means teams can run more streams and allocate compute resources only when something is happening. Developers control the detection sensitivity thresholds to tune their baselines.

Motion detection in VIF runs entirely on CPU, sending an alert when any motion appears in a specific area. Object detection, scene analysis, and synthetic video detection still require a GPU.

How Can Teams Use Motion-Triggered Analysis?

Motion-triggered analysis is suited to workflows with many feeds that sit idle for long periods of time. For public safety, for instance, security staff can be alerted whenever motion appears on cameras that cover prohibited areas. Similarly, election offices can monitor ballot boxes and voting booths after hours. When VIF detects motion, it can trigger an alert and send frames to a VLM to describe the scene. Additional example use cases for motion-triggered analysis include:

• Law Enforcement: Monitor prohibited zones and, upon detecting motion, trigger an alert for dispatch to investigate, along with a VLM to generate a description of what happened.
• Industrial Monitoring and Smart Cities: Monitor construction zones or remote worksites during off hours and, if motion is detected, send an alert to public safety or security personnel to investigate further.
• Departments of Transportation: Monitor busy intersections for motion and trigger a VLM or custom object detection to determine whether an accident occurred.
• Healthcare: Monitor patients’ rooms and upon detecting motion trigger a VLM to describe whether the movement is normal or unexpected (for example, to determine if a patient suffered a fall).

What Was New In VIF 1.1?

VIF 1.1 added analysis ofon-demand video using the same models and configurations VIF uses to analyze live video feeds. Version 1.1 also improved the REST API with named webhook secrets and job retention controls, reorganized the user interface into three tabs, and added a non-containerized Wowza Streaming Engine deployment path alongside the standard Docker Compose deployment. VIF 1.1 and later versions require Wowza Streaming Engine 4.11 or newer. Further details on requirements can be found below.

Does VIF Analyze VOD Files?

VIF can analyze recorded files, as well as live streams. On-demand video files can be analyzed through both the REST API and the user interface. Before version 1.1, VIF ran object detection, scene analysis, VLM analysis, and synthetic video detection against live streams only. Now, users can run the same models and saved configurations against video on-demand (VOD) files.

The same configurations that a team has for live streams can also apply to recorded VOD files. A tuned threshold and detector combination for a live camera feed can be reused against a VOD archive. Each job processes the entire file as fast as the analysis backend allows. Then, VIF produces a complete record containing timestamped detections that can be queried, along with a job manifest.

How Does VIF Handle VOD Analysis?

Every VOD job that stores results generates a corresponding JSONL file. That JSONL file carries the results, metadata, and timestamps for tools like Splunk and Elasticsearch. When a job fails, VIF resumes analysis from the last answered window, after confirming that neither the source file nor the requested configuration has drifted. Administrators can set both the maximum number of jobs VIF retains, 25 by default, and how long it keeps them.

Reviewing an archive previously meant restreaming it or standing up separate batch tooling. With VIF, the same framework, with the same configurations and data channels, handles both live and VOD workflows.

How Can I Use VIF For VOD Analysis?

VOD analysis in VIF fits many workflows, especially where there are videos or archives that already exist. This includes use cases like post-incident review, archive analysis, content verification, and model evaluation.

A public safety or transportation team could pull a recording and run object detection or scene analysis across it to build a timestamped record of what took place. An organization that deploys VIF for live monitoring can apply the same detector configuration to archives containing historical footage. That makes an archive searchable and actionable, just like the live feed. A media team could run synthetic video detection against submitted or acquired footage before it enters the publishing pipeline. For teams comparing configurations, they could run multiple models or detectors against an identical VOD file to assess performance.

What Changed In The VIF User Interface?

There is a new dashboard for on-demand jobs that lists all running and completed jobs, with detection data, content descriptions, and synthetic video analysis in one place. The redesigned user interface organizes VIF into three tabs, each showing live counts.

1. Live Streams consolidates system health, thumbnail previews, and inline stream controls in one place, with detection overlays drawn directly on the frame.
2. On-Demand handles VOD job submission, management, and results specific to each type of analysis.
3. Configs provides a view of Stream Configs, Stream Config Defaults, and On-Demand Configs, with a Clone Config option for spinning up new Stream Configs without rebuilding from scratch.

With the new updates, an analyst or operator who is reviewing completed VOD analysis jobs now only needs to scrub the timeline to see the detections and confidence scores in context. VIF 1.2 also introduces a model chain builder in the UI to compose multi-stage analysis, select execution modes, and set parameters like motion sensitivity.

How Has The VIF REST API Been Updated?

The updated VIF REST API drives every capability the user interface exposes, including starting or queueing VOD analysis jobs, managing configurations, and retrieving results. Two recent additions are focused on operational security and control:

1. Named webhook secrets let an administrator add webhook endpoint credentials in a secrets document. VIF never exposes the values and never writes them into a job manifest, which keeps credentials out of artifacts that routinely get exported and shared.
2. Retention settings cap how many completed jobs VIF stores, and for how long. A high-volume deployment does not accumulate results indefinitely, and teams can set retention limits that support compliance-sensitive architectures while optimizing storage.

Wowza Streaming Engine and VIF are both highly extensible solutions that can be configured programmatically using the REST and Java API. Everything available in the Manager UI, including the model chain builder, is also available through the REST API.

Does VIF Support More Advanced AI Models?

VIF added support for select vision-language models in a recent release, including NVIDIA Nemotron Nano, Google Gemma 3, NVIDIA Cosmos3, and Qwen3. Version 1.1 of VIF improved VLM analysis. VIF can confirm an external VLM server is reachable and report which model it serves to catch a misconfigured endpoint at setup instead of after a job fails. In VIF 1.2, any supported VLM can serve as a stage in a model chain, and developers can set prompts for each stage.

That expansion reinforces and underscores VIF’s underlying approach to model flexibility. Teams can pair VIF with the models that fit their accuracy, latency, and licensing requirements, including custom-trained models for domain-specific detection work.

How Do I Update To VIF 1.2?

VIF 1.2 runs on Wowza Streaming Engine 4.11 or later versions. While Docker Compose remains the standard deployment method, standalone installers are also available. Existing VIF users can update to the latest version at no extra cost, provided they have an active maintenance and support plan. VIF 1.2 lets teams run heavier analysis only when it matters, across the video feeds they already manage. Get in touch for a tailored demonstration, or see us at IBC 2026.

Frequently Asked Questions

What is new in Wowza VIF 1.2?

Wowza Video Intelligence Framework (VIF) 1.2 adds model chaining and motion detection. Model chaining lets one detection trigger another model, such as object detection or a vision-language model (VLM). The VIF Manager adds a model chain builder UI to compose and configure these chains easily. Motion detection analyzes video feeds for any movement and runs on CPU, without any GPU requirements. Motion detection can serve as its own detection mode, or as part of a model chain. Existing VIF 1.1 configurations continue to work unchanged.

What is model chaining in Wowza VIF?

Model chaining in Wowza VIF links analysis stages so the result of one stage is used to determine what happens next. For example, motion detection can trigger object detection, and an object detection result can trigger a vision-language model to describe the scene, provided the user-defined criteria have been met at all stages of the chain. Chains support three execution modes: Linear, Conditional, and Dynamic. Any supported VLM, including Qwen3-VL, NVIDIA Nemotron, and NVIDIA Cosmos reasoning models, can serve as a stage in a model chain.

Does VIF motion detection require a GPU?

No, motion detection in Wowza VIF runs on CPU and does not require a GPU. Other VIF analysis types, including object detection, scene analysis, and synthetic video detection, require an NVIDIA GPU.

Can VIF send an alert based on motion alone?

Yes, VIF can send an alert, such as a webhook, whenever motion appears in a feed, without running object detection or a vision-language model.

What is new in Wowza VIF 1.1?

Wowza’s Video Intelligence Framework (VIF) version 1.1 added four major capabilities: analysis of recorded video on-demand (VOD) files, an updated REST API, a redesigned Manager interface, and a non-containerized Wowza Streaming Engine deployment path. VIF 1.1 also expanded support for vision-language models (VLMs) and added verification for external endpoints.

Does VIF analyze VOD files?

Yes, VIF analyzes VOD files as well as live streams. On-demand analysis works through both the REST API and Wowza Streaming Engine Manager, supporting object detection, scene analysis, VLM analysis, and synthetic video detection workflows.

What Wowza Streaming Engine version does VIF require?

Wowza VIF 1.2 requires Wowza Streaming Engine 4.11 or later. VIF can be deployed via the standard Docker Compose path or installed with a standalone installer. Existing VIF 1.1 configurations continue to work without changes.

What are the tabs in the VIF Manager interface?

The VIF Manager interface contains three tabs: Live Streams, On-Demand, and Configs. Each of these tabs contains live counts and information about analysis jobs. Live Streams shows system health, thumbnail previews with detection overlays, and inline stream controls. On-Demand shows job status and results for VOD projects. Configs provides a view of Stream Configs, Stream Config Defaults, and On-Demand Configs.

What is in the VIF REST API?

The updated VIF REST API lets a user submit an on-demand job, manage configurations, and retrieve results. It also supports named webhook secrets and retention controls. In VIF 1.2, the REST API exposes everything available in the Manager UI, including the ability to build and configure model chains, as well as motion sensitivity settings.

Which vision-language models does Wowza VIF support?

VIF supports various vision-language models (VLMs), including NVIDIA Nemotron Nano 12B VL, Google Gemma 3 4B, NVIDIA Cosmos3 Edge, NVIDIA Cosmos3 Nano, and Qwen3-VL-4B. These are selectable in VIF, but selecting a model does not deploy it automatically. The matching VLM sidecar must be running as well.

How does VIF handle a failed on-demand analysis job?

For VOD, VIF resumes a failed on-demand job from the last answered window, only after confirming that neither the source file nor the requested configuration has drifted. VIF will retry a failed job up to three times in an effort to complete it satisfactorily. This process resets whenever a retry attempt makes progress past the previous run, so a long file with occasional interruptions is not stalled and can keep advancing.

Can VIF detect synthetic video in recorded files?

Yes, VIF can detect synthetic video content in recorded VOD files. VIF can run the NVIDIA Synthetic Video Detector and return a per-window confidence score against a configurable threshold across both live and on-demand workflows. Currently, model chains only support SVD as a stage for on-demand (VOD) analysis.

Did Wowza win a Streaming Media award in 2026?

Yes, Wowza Streaming Engine won the 2026 Streaming Media Readers’ Choice Award for Real-Time Streaming Solution. VIF was also named one of three finalists for Streaming Innovation of the Year.