Video compressed for streaming can lose fine details before it reaches a viewer’s screen. Nokia wants the device playing it back to use AI to put some of that quality back.
Nokia researchers have released an open-source implementation of Neural Network Post Filter technology, or NNPF, which applies neural-network processing after a compatible device decodes a video. Nokia says the work marks the first time an AI-based video-enhancement method has been integrated into international video standards.
The approach could eventually help streaming services improve perceived video quality without abandoning established codecs. But realizing those benefits will depend on support from encoding platforms, chipmakers, device manufacturers, and video applications.
How Nokia’s neural-network filter works
Video compression reduces the amount of data required to transmit and store content, but it can also remove fine visual information. NNPF applies a neural-network filter after a device decodes the compressed video, allowing the playback system to enhance the resulting image and recover some of the detail lost during compression.
The system uses supplemental enhancement information, or SEI, carried alongside the coded video. This metadata gives a compatible receiving device the information needed to identify and apply the appropriate neural-network filter.
Nokia researchers Miska Hannuksela and Jill Boyce said the company contributed to the Versatile Supplemental Enhancement Information standard, which defines ways to carry this type of metadata with coded video.
Because the metadata supplements established codecs such as High Efficiency Video Coding (HEVC) and Versatile Video Coding (VVC), providers may be able to introduce AI enhancement without moving to an entirely new compression format.
That does not make deployment automatic. Video services would still need compatible encoding tools, while playback devices and applications would need software capable of interpreting the metadata and running the neural network.
AI enhancement without a high-end GPU
Other players in the tech space have also been working on AI-based video enhancement. However, their solutions have either been costly or highly restrictive.
NVIDIA provides the GeForce RTX GPU for AI-enabled video enhancement, fixing compression flaws, and SDR to HDR tonemapping. Despite all these features that NVIDIA provides, its hardware comes at a price higher than most consumer devices. Additionally, their developer tooling is restrictive and makes it difficult for collaboration.
To elaborate, the NVIDIA RTX software development kit (SDK) license terms read: “You may not use the SDK in any manner that would cause it to become subject to an open source software license.” The message regarding developer freedom is clear.
Google also has AI video enhancements; though, it has less features compared to NNPF. Google's optimizations are mainly featured around resizing content and voiceover. Additionally, Google restricts some features like voiceover to only marketing campaigns under their Performance Max tier.
What this could mean for video providers
Consumer devices already have native AI acceleration for tasks like deep learning. Moreover, the ecosystem for support of open technologies for local AI inference is not just by Nokia.
Intel also has an open-source tool called OpenVINO, which supports CPUs, GPUs, and NPUs. This goes to signal support for public goods, without merely looking at corporate interests.
As enterprise teams rush to optimize their stack on the backend, they might need to consider the already existing hardware solutions that their users interact with on everyday devices. They could see lower video CDN costs for their businesses and greater user satisfaction through immersive video experience.
Also read: AI Videos Let South Korean Families See Dead Relatives Again to explore how AI video technology is reshaping family remembrance—and raising new ethical concerns.