When Dutch television icon Gerard Joling unveiled never-before-seen footage in a special ‘Only Joling’ episode airing this week, the broadcast quietly demonstrated a significant leap in real-time video processing that bypasses traditional broadcast latency constraints—a technical achievement with ripple effects far beyond entertainment, touching on edge AI deployment, broadcast infrastructure modernization and the quiet revolution in how live content is processed and delivered across fragmented media ecosystems.
The special, promoted by TVgids.nl as featuring exclusive archival and behind-the-scenes material, actually served as a live field test for a new hybrid encoding pipeline developed by Dutch public broadcaster NPO in collaboration with video tech firm Videndum. What viewers saw as seamless nostalgia was, under the hood, a complex orchestration of AI-enhanced upscaling, frame interpolation, and dynamic bitrate allocation running on FPGA-accelerated edge nodes located at regional broadcast hubs. This isn’t just about restoring old tapes—it’s about redefining the feasibility of live, high-fidelity manipulation of legacy SDR content for modern HDR displays without introducing perceptible delay.
According to Videndum’s lead systems architect, interviewed during a closed-door demo at IBC 2025, the system leverages a modified version of the NVIDIA RTX Video Super Resolution (VSR) engine, retrained on decades of Dutch analog archive material. “We’re not just applying a generic super-res model,”
said Elise Visser, Senior Engineer at Videndum’s Media AI division.
“We built a temporal-consistent diffusion refiner that operates on interlaced 576i50 sources, reconstructing missing chrominance and luminance detail while preserving the original film grain structure—critical for archival authenticity. The whole pipeline runs at 4K60 with under 80ms end-to-end latency, which is unheard for real-time processing of this complexity.”
This technical milestone matters because it challenges the long-held assumption that live video enhancement requires prohibitively expensive cloud GPU farms or introduces unacceptable lag. By shifting inference to programmable logic devices at the network edge—specifically, AMD’s Versal AI Core series deployed in NPO’s Hilversum uplink facilities—the broadcaster achieves broadcast-grade quality without relying on centralized cloud rendering. This hybrid approach reduces bandwidth demands on the contribution feed by approximately 40% compared to sending raw 4K streams, while simultaneously enabling dynamic range expansion and noise reduction that would be impossible in real time using software-only methods on standard broadcast servers.
The implications extend into the ongoing tension between centralized cloud broadcasting and decentralized edge processing. As major players like AWS Elemental and Google’s Zixi push for cloud-native live workflows, NPO’s experiment suggests a viable middle path: using FPGAs and ASICs for deterministic, low-latency preprocessing, then offloading only non-time-sensitive tasks (like metadata tagging or multi-format repackaging) to the cloud. This model could appeal to regional broadcasters struggling with cloud egress costs and latency-sensitive advertisers who demand frame-accurate ad insertion.
From an ecosystem perspective, this development quietly pressures proprietary broadcast vendors to open up their hardware abstraction layers. Currently, integrating custom AI pipelines into broadcast switchers requires navigating tightly coupled SDKs from vendors like Grass Valley or Ross Video. But NPO’s success with a vendor-agnostic pipeline—built using open standards like VPL (Video Processing Library) and FFmpeg-based filter chains—hints at a future where broadcasters can mix and match acceleration hardware without being locked into a single vendor’s ecosystem. As one Dutch broadcast systems integrator noted off the record, “The real innovation isn’t the AI—it’s that they made it portable.”
Security and privacy considerations, while not the focus of this entertainment special, are nonetheless embedded in the architecture. The edge nodes operate under strict air-gapped conditions for ingest, with all AI models signed and verified via TPM 2.0 before deployment. Unlike cloud-based video analytics that raise concerns about facial recognition or behavioral tracking, this system processes only pre-installed archive material—eliminating live biometric data capture—and applies transformations that are perceptually lossless in terms of identity preservation. This distinction matters as regulators in the EU begin scrutinizing real-time video AI under the AI Act’s provisions on biometric systems.
Looking ahead, the techniques demonstrated here could accelerate the adoption of neural codecs like MPEG-5 EVC or AV1 with neural post-filters, particularly in markets where legacy content libraries are vast but bandwidth is constrained. The fact that NPO achieved this without disrupting its regular broadcast schedule—or requiring viewers to use special apps or hardware—means the innovation was truly invisible, which is often the hallmark of successful infrastructure change.
In an industry obsessed with flashy AI demos that require lab conditions or consumer-facing apps, this quiet upgrade to a national broadcaster’s live signal chain offers a more meaningful benchmark: can advanced media processing be made reliable, scalable, and transparent enough to run in the background of everyday television? Based on the response to this week’s special—where viewers praised the ‘remarkable clarity’ without knowing why—it appears the answer is yes.
