Guides / AWS Elemental Inference
AWS Elemental Inference: quick answers
Short answers about what AWS Elemental Inference outputs, which of its signals reach the ad request at the break, and where Sceney, the vision language model (VLM) from Edge Video AI, works alongside it. Updated November 2026.
Sceney is the vision language model (VLM) from Edge Video AI. It watches live video, reads what's on screen and what's said, and turns it into signals ads and shopping can act on in the moment.
Works alongside your existing playout and ad server.
Quick answers
- What does AWS Elemental Inference output?
- AWS lists four built-in features: vertical video (smart crop), highlight clips (event clipping, for soccer and basketball matches today), same-language smart subtitles in TTML, and contextual metadata that classifies content at shot and scene level, including IAB content categories, brand safety signals, objects, actions and scene descriptions. Customers can add their own yes/no or pick-one labels (decisions) and run their own prompts on an Amazon Bedrock model they choose (functions). See What is Elemental Inference? in the AWS user guide.
- Does AWS Elemental Inference choose the ads?
- No. In the ad flow AWS documents, Elemental Inference supplies the classifications, MediaTailor passes them on, and your ad decision server uses them to target ads. Sceney, the vision language model (VLM) from Edge Video AI, works the same way: it adds context to the request and leaves the choice of ad to the ad server.
- Which Elemental Inference fields reach the ad request at the break?
- In AWS's MediaTailor guide, the contextual ad targeting setup passes IAB content categories and GARM brand safety signals. They are written to keys such as player_params.iabCategories, which you can rename to match your ad decision server's parameters. AWS says its recipe uses only the IAB and GARM fields, and suggests turning off summary generation and extended analysis when contextual metadata is used only for ad targeting, because those descriptive fields add processing time before the break. See Contextual ad targeting with Elemental Inference.
- What part of the programme does the break query cover?
- According to AWS, MediaTailor asks Elemental Inference for the content that played since the end of the previous ad break, up to the longest lookback Elemental Inference supports, so the categories describe what aired just before this break rather than the last ad pod. Sceney also holds the context from just before each break.
- What happens if the contextual data isn't ready at the break?
- AWS says MediaTailor does not wait for processing that is still in progress, and that if the function fails (a timeout, an authorization error or an empty response), ad insertion goes ahead without it and the break is not blocked. Sceney follows the same rule: if the context isn't ready, the request goes out as it always would. Edge can add information; it cannot block a request.
- Does AWS Elemental Inference read speech and on-screen text?
- AWS says smart subtitles use automatic speech recognition to make same-language subtitles, and that the extended analysis setting on contextual metadata detects people, environments, brands and on-screen text. Sceney reads what's on screen and what's said, including speech and the text and logos in the picture, and turns that into IAB categories, scene keywords and brands seen on screen. See Speech and on-screen text for ad requests.
- Can AWS Elemental Inference spot brands on screen?
- AWS lists brands among the descriptive fields that extended analysis detects, and gives measuring sponsor logo exposure as an example of a custom function run on an Amazon Bedrock model you choose. Brands seen on screen are part of what Sceney returns, next to IAB categories and scene keywords.
- Who is AWS Elemental Inference for if you're not on MediaLive?
- AWS says Elemental Inference is designed for any customer with a video workload, and that it is set up through AWS Elemental MediaLive and MediaConvert. The MediaTailor ad targeting flow AWS documents uses a MediaLive channel with contextual metadata enrichment turned on to mark each break; without that, AWS says you must work out the timing values from your own pipeline. If your playout and ad server are elsewhere, Sceney reads the live stream itself and its fields can be added to your existing ad request. Works alongside your existing playout and ad server.
- Can Sceney's IAB categories sit next to Elemental Inference signals in one ad request?
- An ad request can carry more than one source. AWS's MediaTailor guide shows the Elemental Inference call chained with other functions before the request goes to the ad decision server. Sceney's IAB categories, scene keywords and brands seen on screen are added to the publisher's existing request by its ad server or SSAI calling Edge at the break, next to what is already there, not instead of it. See IAB categories in live ad requests.
- Is Edge Video AI affiliated with AWS?
- No. Edge Video AI is not affiliated with Amazon Web Services. Sceney reads the live stream on its own. It does not need Elemental Inference output to work. The AWS facts here come from AWS's own public pages, listed under Sources. Edge Video AI makes Sceney; it is based in The Hague and was founded in 2019. See also AWS Elemental Inference and Sceney.
A fair distinction
We describe AWS Elemental Inference from AWS's own public pages, checked in November 2026 and listed under Sources below. We do not claim a complete description of the service, and we do not speak for AWS. Sceney works alongside Elemental Inference; it is not a replacement for it. Sceney reads the live stream on its own. It does not need Elemental Inference output to work.
Related
AWS Elemental Inference and Sceney
Who does what, side by side, in an AWS Elemental workflow.
Sceney
The vision language model that reads what's on screen and what's said.
IAB categories in live ad requests
How those categories reach the request you already send.
Anoki ContextIQ: quick answers
Scene-level context for CTV ads, and where Sceney fits.
Sources
- AWS, AWS Elemental Inference, product page, read 20 Nov 2026. Source for: AI applied in parallel with encoding; vertical video, clip generation, smart subtitles and contextual metadata (IAB content categories, brand safety signals, objects, actions, scene descriptions); set up through MediaLive and MediaConvert.
- AWS, AWS Elemental Inference FAQs, read 20 Nov 2026. Source for: who it is designed for; clip events for soccer and basketball matches today; decisions (your own yes/no or pick-one labels) and functions (your own prompts on an Amazon Bedrock model you choose), with sponsor logo exposure and on-screen graphics given as function examples.
- AWS, What is Elemental Inference?, AWS Elemental Inference User Guide, read 20 Nov 2026. Source for: smart crop, event clipping, smart subtitles using automatic speech recognition (TTML), contextual metadata at shot and scene level, customer decisions and functions.
- AWS, Contextual ad targeting with Elemental Inference, AWS Elemental MediaTailor User Guide, read 20 Nov 2026. Source for: IAB and GARM signals passed to the ad decision server, which targets the ads; output keys such as player_params.iabCategories; the query window since the previous break; summary generation and extended analysis (people, environments, brands, on-screen text) adding processing time; the MediaLive enrichment prerequisite; chaining with other functions; ad insertion going ahead if the function fails.
AWS, AWS Elemental, MediaLive, MediaConvert, MediaTailor and Amazon Bedrock are trademarks of Amazon.com, Inc. or its affiliates. Edge Video AI is not affiliated with Amazon Web Services.
Running AWS Elemental? Add the moment.
Works alongside your existing playout and ad server.