Live sport played in sun and shade is one of the hardest lighting situations a broadcast has to handle. Half the pitch is in direct sunlight, the other half in deep shadow, and the camera cannot expose for both at once. We covered the picture problem in shadows in live sports. This article is about the other side of it: what it means to put real-time AI video processing into an outside broadcast chain, and the engineering questions an OB team should ask before a trial.
LivePictureAI is a real-time AI processing layer that balances sunlight and shadows in live video. It sits either on each camera feed before the vision mixer or on the programme output after it. That choice of placement shapes almost every decision below.
The latency budget of a live chain
Every live chain has a latency budget, whether or not anyone has written it down. Each stage between the lens and the viewer takes its share: camera processing, transport, frame synchronisation, the vision mixer, graphics, encoding and distribution. Some stages are fixed by the hardware; others depend on configuration. What matters to the production is the total, and whether the pictures stay in step with audio, commentary, replays and any other feeds that have to line up.
An AI stage is just another entry in that budget. LivePictureAI adds at most 100 ms of latency. That figure is the whole cost of the stage, and it is the number to plug into your own calculation. Whether it is comfortable depends on your chain: how much headroom you already have, how audio is delayed to match, and whether the processed feeds need to stay aligned with feeds that are not processed.
The practical point is to treat the AI stage like any other piece of equipment in the truck. Measure the chain before it goes in, measure it after, and make sure audio and any unprocessed sources are compensated accordingly.
Before the vision mixer or after
LivePictureAI supports two setups, and the right one depends on the size of the production.
Large productions
Smaller productions
Before the vision mixer, the processing works on each camera feed individually. The director, vision engineers and replay operators all work with pictures that have already been balanced. This suits large productions, and it needs more compute: a GPU per camera feed.
After the vision mixer, a single process works on the programme output. There is one stream to process, so the compute footprint is lighter. This suits smaller productions, where the cut is made as today and the finished programme is enhanced on its way out.
Either way the roles stay clear. Vision engineers keep camera control: iris, shading, camera matching and the overall look remain their decisions. The AI handles the sun-and-shadow balance inside the picture, so there is less manual chasing of the shadow line as it moves across the pitch during a match.
Signal format follows from the same placement. The layer has to accept what the cameras and routing already deliver. LivePictureAI works with SMPTE 2110, NDI, SDI, RTSP and file-based streams, so it can be inserted into an IP facility or a baseband truck without changing the cameras. For 2110 chains, timing comes first; ourPTP locking guide covers the basics.
GPU edge inference for broadcast video
Real-time inference on live video is a different problem from processing a file. A file can be processed at whatever speed the hardware allows; a live feed has to be processed at the frame rate, every frame, with a predictable delay. That is why the processing runs on GPUs close to the signal.
LivePictureAI is designed for GPU-based deployment at the edge or in the production hub. In practice that can mean hardware in or next to the OB truck at the venue, or in a central facility where feeds are brought back for remote production. It can run on-premises or in the cloud, depending on where the production already does its processing and how the feeds are transported.
The choice between venue and hub is usually made by the production model rather than by the AI. A traditional OB production keeps processing in the truck; a remote or REMI production may prefer to process at the hub, where the vision mixer is. The enhancement layer should follow the vision mixer, not the other way round.
Sizing, redundancy and fallback: the questions to ask
Hardware sizing and resilience depend on the production, so they are best approached as evaluation questions rather than a checklist of features. When assessing any real-time AI processing stage for a live chain, it is worth asking:
- How many feeds need processing? All cameras, or only those covering the parts of the pitch where sun and shadow meet?
- Where will the hardware live? In the truck, at the venue, at the hub or in the cloud, and what does that mean for power, cooling, rack space and transport?
- What is the latency of the full chain with the stage in place? Measured end to end, not only for the stage on its own.
- What happens to a feed if the stage stops? How bypass and fallback are designed, and who decides to use them during a live event.
- How does the stage fit existing redundancy? Whether it follows the facility's current approach to main and backup paths.
- Who operates it? What the vision engineers see, what they can adjust, and what stays fully under their control.
These answers vary between productions and facilities, and they are part of the integration discussion. We are validating LivePictureAI with selected partners after IBC 2026, and these are exactly the questions those evaluations are built around.
Monitoring
An engineer will only trust a processing stage they can see. Before going live, it should be possible to compare the source feed and the processed feed side by side on the monitoring wall, so the vision engineers can judge the correction against what the camera is delivering. During the production, the processed feeds should be watched like any other source: picture, timing and signal health.
Monitoring also helps build confidence over a trial. Watching how the correction behaves as the shadow line moves across the pitch, as clouds pass and as the light drops into evening tells you more than any specification. If you are planning a trial and want to talk through where the layer would sit in your chain, seeLivePictureAI for broadcasters or book a demo.