Scene-Level AI Targets Live CTV Brand Safety Checks

By Amara Osei

89% of advertising requests analysed in Wurl’s October 2026 CTV Trends Report lacked a usable programme name.

Another 45% lacked a usable channel name, while 35% of analysed inventory carried a declared genre that did not match the programme. Wurl counted more than 3,600 different genre labels in a two-hour prime-time window and found labels fragmented, vague or inconsistent across services.

As connected television (CTV) advertising expands, audience data shows who may be watching, but sparse content data leaves buyers guessing what appears beside an ad, making contextual targeting (ad placement based on surrounding content) harder.

What can scene-level AI actually see?

Scene-level artificial intelligence (AI) analyses the video itself, scene by scene, rather than treating programme descriptions, channel classifications or metadata (descriptive tags) as the whole truth.

A multimodal AI system (one that combines several data types) reads visuals, dialogue, background sound, subtitles and on-screen text; an object or keyword alone cannot establish meaning.

A person holding a weapon, for example, may appear in fiction, a news report or a documentary, so audio and surrounding context can change the safety judgement.

Platforms can apply these signals to brand-safety policies, which exclude graphic violence, hate speech and explicit content, or to brand suitability, which applies an advertiser’s values, audience and campaign goals.

The distinction matters because a news broadcast can move from conflict coverage to financial results and a human-interest story within minutes; one label for the entire programme wastes suitable inventory.

The same analysis can support content moderation, video validation and media intelligence (structured information extracted from video).

Live use is harder: video on demand can be checked before release, while live news and sports can change subject within seconds, forcing systems to balance speed and accuracy and decide when a scene needs another pass.

Signals also have to reach advertising platforms and buyers quickly and consistently; a scene label that arrives after the ad decision cannot influence placement.

Scene analysis offers finer context, but only when it delivers a usable answer in time.

What is PYLER testing in live CTV?

South Korean AI company PYLER is developing a multimodal engine for digital video and CTV, reading visual, audio and textual signals to assess safety and advertising suitability.

At the Asia-Pacific finals of the NVIDIA Inception Grand Challenge 2026, PYLER presented a project with a global free ad-supported streaming television (FAST) provider and reached the Top 10.

It examines dialogue, on-screen text, sound, brands, locations and actions in individual live CTV scenes, including when conventional signals cannot separate programming from advertisements.

PYLER CEO Jaeho Oh said:

“Live video presents a different challenge from pre-recorded content. Scenes change continuously, and systems need to interpret what is happening without relying entirely on predefined metadata or program signals. PYLER addresses this challenge through multimodal video understanding and efficient, context-aware classification.

Our AI model analyzes visual, audio, and textual signals to interpret scenes and evaluate them against different safety and suitability criteria.

For live applications, we are developing more efficient, probability-based approaches that can adapt to different evaluation criteria. When a result has low confidence or requires further analysis, additional models can be used rather than relying on a single classification decision.

This approach is designed to balance speed, accuracy, and computational efficiency. It is particularly important for unpredictable content, such as breaking news, where the meaning of a scene can change rapidly.”

PYLER has presented the workflow; its live deployment performance remains under evaluation.

What do the performance numbers prove?

PYLER’s NVIDIA collaboration underpins the work. A June 2026 NVIDIA customer case study reports:

• Two million videos processed daily.
• Unsafe-content exposure reduced by 76.7% for major advertisers.
• AI validation time of 0.011 seconds, against a 30-second human-review benchmark.
• A tenfold reduction in operational costs.

Oh said:

“Our NVIDIA-powered infrastructure has already demonstrated significant efficiency gains. According to NVIDIA’s published case study in June, PYLER reduced content validation time from a 30-second human-review benchmark to 0.011 seconds for AI validation. The same case study reported a tenfold reduction in operational costs.

These figures reflect our broader video validation infrastructure rather than the end-to-end latency of a live streaming deployment.

We are continuing to evaluate live-specific performance, including processing latency, classification accuracy, and the handling of ambiguous scenes. Our goal is to provide reliable, context-aware intelligence that allows platforms to apply their own standards with greater speed and consistency.”

The figures show validation throughput, not end-to-end live-streaming latency.

What still has to work in the ad stack?

More precise signals could reduce exclusions of whole programmes or genres when individual segments fit a campaign. They could help CTV and FAST operators sell inventory, give advertisers relevant placements and show where campaigns appear.

Operational plumbing is moving too. On October 28, 2026, IAB Tech Lab released its Live Event Ad Playbook (LEAP) implementation guidance for public comment. It aims to improve coordination among buyers, publishers and advertising-technology providers through standardised forecasting and near-real-time audience signals.

LEAP focuses on delivery and operational coordination, not scene-level analysis, but addresses part of the scaling problem. PYLER still needs low-latency processing, reliable interpretation and integration with programmatic advertising systems (automated ad buying) before buyers can use scene intelligence at scale.

Scene intelligence will matter only when it changes real ad decisions, quickly and repeatably.