Live outdoor sport is shot in light that no camera can fully tame. Bright sunlight on one half of the pitch and deep shadow on the other force a compromise on every frame. LivePictureAI uses real-time AI to balance those highlights and shadows on the live feed itself, the problem we describe in more detail inshadows in live sports. This page covers the practical side: how that enhancement layer is added to an existing NDI or SDI workflow without replacing the equipment around it.
The short version: you keep your cameras, your switcher and your people. LivePictureAI is inserted on the camera feeds before the vision mixer or on the programme output, takes the signal in over the transport you already use, and hands a corrected feed back to the rest of the chain.
Where the enhancement layer sits in an NDI or SDI chain
There are two setups, depending on the size of the production:
Large productions
Smaller productions
In a large production the layer sits on each camera feed before the vision mixer, so the director's cut, replays and graphics all work with pictures that have already been balanced. That needs more compute, a GPU per feed. In a smaller production it sits on the programme output after the vision mixer: one stream to process, and a lighter footprint. Either way, the next device does not need to know the layer is there: it receives a feed over the same transport it expects.
Camera control stays where it is. Vision engineers keep their racks and remain responsible for camera matching, exposure decisions and overall picture quality. The aim is to take the constant sun-and-shadow balancing off their hands, not to take the cameras away from them.
Supported transports: SMPTE 2110, NDI, SDI, RTSP and file
Broadcast facilities rarely run a single transport. A venue might deliver SDI from the cameras, a production hub might run SMPTE 2110 internally, and a smaller or remote production might live entirely on NDI. LivePictureAI is designed to fit into whichever of these you use:
- SDI: the established baseband connection in trucks and at venues.
- NDI: IP video over standard networks, common in remote and software-based production.
- SMPTE 2110: uncompressed IP video in larger facilities and production hubs. If you work with 2110 yourself, our 2110 Toolkit helps you analyse and generate streams.
- RTSP: for network camera and contribution sources.
- File-based streams: useful for evaluating the enhancement on your own recorded footage.
The enhancement itself is the same whichever transport carries the picture. What changes between setups is the plumbing around it, which is why integration details are agreed per installation rather than assumed.
Latency budget
Every element in a live chain spends some of the latency budget, and an AI layer is no exception. LivePictureAI adds at most 100 ms of latency, built to keep pace with live broadcast timing.
How that sits within your overall budget depends on everything else in the chain: the cameras, any conversion, the switcher, graphics and the encoding for distribution. With the layer on the camera feeds, the added delay applies equally to every feed going into the vision mixer, which keeps their timing consistent for the director. On the programme output, it applies once, to the finished mix. We look at the full chain together with you, so the number is understood in context rather than in isolation.
At most 100 ms added. One figure for the processing layer; how it fits your end-to-end budget is worked out per setup.
Deployment: GPU edge, production hub, on-prem or cloud
LivePictureAI runs on GPU infrastructure, and it is designed to be deployed where your feeds already are:
- GPU-based edge: close to the cameras, at the venue or in the OB truck, which suits SDI-heavy productions. We cover the truck case in low-latency AI processing in an OB truck.
- Production hub: centralised processing for remote and IP-based productions, where NDI and SMPTE 2110 are common.
- On-prem or cloud: on a local GPU, or in the cloud on AWS or Azure.
Which option fits best depends on where the switcher lives, how the feeds travel and what infrastructure you already operate. None of them requires new cameras.
SDI low-latency AI processing
SDI is still the backbone of many outside broadcasts, and it is where low latency matters most to operators used to baseband timing. For SDI chains, the enhancement layer is typically deployed at the GPU edge (in the truck or at the venue), so the feeds are processed close to the cameras and handed straight on to the vision mixer, or the programme output is processed on its way out.
From the vision mixer's point of view, nothing changes about how it receives the camera feeds; in the large production setup they simply arrive already balanced for sun and shadow. The same principle applies to NDI: the layer receives the feed, corrects it in real time, and passes it on, so the rest of the network carries on as before.
What we discuss in an evaluation
Integration requirements depend on your existing infrastructure, so we do not publish a one-size-fits-all specification. Formats, resolutions and frame rates are agreed per setup. In an evaluation we typically go through:
- Which transports your feeds use today (SDI, NDI, SMPTE 2110, RTSP or a mix) and where they converge.
- The signal formats your production runs, agreed for your specific setup.
- Your end-to-end latency budget and where the processing layer fits within it.
- Where the GPU processing should run: edge, production hub, on-prem or cloud.
- Which setup fits (camera feeds or programme output), how the corrected feed is handed on, and how your vision engineers work alongside it.
- Test material: ideally your own footage, which can be run through as file-based streams.
We are validating LivePictureAI with selected partners after IBC 2026. If you run NDI or SDI productions in outdoor sport and want to see the enhancement on your own feeds, see how it fitsbroadcast and OB workflows or get in touch below.