Media and entertainment
How studios, broadcasters and content producers use AI and NVIDIA technologies: ranking and enhancing live camera feeds at the venue, localizing programs into other languages, tuning models on their own sports footage, and shared 3D previs tools, with notes on when existing production tools suffice.
The problem
Live production is unforgiving: a mistake goes to air the moment it happens. A single sports event can involve dozens of cameras, and directors decide in seconds which feed matters. Outside broadcasts haul racks of servers in trucks, venue radio links get congested, and sending contribution video to a distant cloud adds delay and exposes rights-protected footage.
Audiences are spread across languages and platforms, yet subtitling, dubbing and regional versions are still largely manual, which limits how much of a catalog or a live schedule can be offered outside the home market. Archives hold decades of footage that nobody can search properly because the metadata was never written.
Studios face a different squeeze. Directors and story artists want to explore camera moves and staging early, but many of them do not work in 3D tools, and pushing scene files between departments can take days. Every round of review that waits on a file transfer is a round that does not happen.12
The approach
For live production the pattern is to process feeds close to the cameras. FanDuel TV's pilot with Verizon placed GPU servers with AI Enterprise and Holoscan for Media at the track; video from a simulated race reached them over portable private 5G, models ranked the cameras, and a language model located one named jockey by the color of the silks. NVIDIA describes Holoscan for Media as an open reference architecture and toolkit for live media; it is listed here separately from the Holoscan SDK.
For localization and enhancement, NVIDIA AI for Media packages NIM microservices such as LipSync, Active Speaker Detection and Studio Voice, and a Content Localization reference workflow inside Holoscan for Media covers captions, translated audio, dubbing and localized graphics. Its Sports Intelligence Playbooks describe fine-tuning Nemotron and other open models on a league's own footage and annotations. On the studio side, Sony Pictures Animation built FlixiVerse, a browser-based previs tool, on Omniverse Enterprise and OpenUSD and used it in pre-production of Spider-Man: Across the Spider-Verse.
Much of this can be done without new infrastructure. Existing editing suites and media asset managers already offer transcription and automatic tagging, cloud captioning and dubbing services handle recorded content at moderate volume, and a fixed studio with cabled links rarely needs private 5G or edge GPU servers. Custom edge inference pays off where latency, rights or connectivity rule out the cloud, or where volumes make per-minute services expensive.12345
Conceptual architecture
Diagram as a list
Applications & solutions
- Production switcher, multiviewer and playoutKeep editors in control and send the program to air or streamingConnects to Media asset management and archive
- Media asset management and archiveStore programs with generated metadata for later searchConnects to Models tuned on the rights holder's own footage (Nemotron)
- Studio previs and layout tools (Omniverse on OpenUSD)Let story and layout artists stage shots in shared 3D scenesConnects to Media asset management and archive
Models & frameworks
- Feed ranking and detection modelsScore camera feeds, detect players, riders or logos in real timeConnects to Production switcher, multiviewer and playout
- Models tuned on the rights holder's own footage (Nemotron)Recognize sport-specific moments for highlights and searchConnects to Venue or facility GPU servers (AI Enterprise, Holoscan for Media)
Inference & runtime software
- Localization services (NIM: LipSync, Active Speaker Detection, Studio Voice)Produce captions, translated speech and matched lip movementConnects to Production switcher, multiviewer and playout
Accelerated computing
- Venue or facility GPU servers (AI Enterprise, Holoscan for Media)Run media functions and AI models next to the feedsConnects to Feed ranking and detection models, Localization services (NIM: LipSync, Active Speaker Detection, Studio Voice)
Networking, power & facilities
- Cameras, microphones and contribution links (ST 2110, private 5G)Carry live video and audio from the venue to processingConnects to Venue or facility GPU servers (AI Enterprise, Holoscan for Media)
Technologies and their roles
NVIDIA AI Enterprise1
Supported software layer for live media AI
Verizon's broadcast framework, tested in the FanDuel TV pilot, is built on NVIDIA AI Enterprise, the supported suite that NVIDIA says includes NIM microservices.
NVIDIA NIM5
Localization and enhancement microservices
NVIDIA AI for Media offers NIM microservices for lip sync, active speaker detection and studio-quality voice, used in its Content Localization workflow.
NVIDIA Nemotron5
Open models tuned on sports footage
NVIDIA's Sports Intelligence Playbooks describe fine-tuning Nemotron and other open models on a sport's own footage and annotations.
NVIDIA Omniverse2
Shared 3D previs pipelines
Sony Pictures Animation built its FlixiVerse previs app on Omniverse Enterprise and OpenUSD for pre-production work.
What you need first
- An IP production chain (SMPTE ST 2110 or a similar standard) or a plan to bridge baseband signals into one
- Rights and contract terms that allow footage, voices and likenesses to be processed and, where planned, used to tune models
- Labeled examples of the moments, players or objects the models should recognize, drawn from your own archive
- Editors and directors who define when automation may act on air and when a person must approve
- Engineering staff or an integrator able to run containers and Kubernetes on GPU servers at venues or in the facility
- Measured baselines for latency, staffing and turnaround so pilot results can be compared
Risks and how to reduce them
- Errors going straight to air
- Keep a human in the loop for anything that changes the program output, start with assistive functions such as feed ranking or metadata, and keep a manual fallback path in the control room.
- Consent for synthetic voices and lip movement
- Get written consent from performers and presenters for dubbing and lip sync, follow union and talent agreements, and label altered content where rules or audience trust require it.
- Personal data in live feeds and archives
- Crowds and bystanders appear in footage; limit face recognition to agreed purposes, apply retention rules to generated metadata, and check data protection law in each market.
- Unverified vendor results
- Published figures for live media AI come mostly from vendors and pilots with simulated events; run your own trial on real programming before changing staffing or contracts.
Documented examples
FanDuel TV · Media and entertainment (live sports broadcasting)
FanDuel TV: piloting AI-assisted live race production on NVIDIA AI Enterprise
FanDuel TV, which televises horse racing on three linear networks, piloted AI-assisted live production with Verizon and partners. Feeds from a simulated race crossed a portable private 5G network to edge servers running NVIDIA AI Enterprise and Holoscan for Media, where models ranked them. Shown at NAB 2025.
Pilot
Related
Sources
- Verizon Private 5G Network with Enterprise AI for live media (customer story) (opens in a new tab)
- Sony Pictures Animation: bringing 2D and 3D worlds together with Omniverse Enterprise (customer story) (opens in a new tab)
- Verizon Business debuts broadcast private 5G and AI solution at NAB 2025 (opens in a new tab)
- NVIDIA Holoscan for Media (developer page) (opens in a new tab)
- NVIDIA brings real-time AI to broadcast, sports and global streaming at IBC (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.