Google Patents an AI Filter That Sharpens Blocky Streaming Video Without Draining Your Processor
Every compressed video you watch carries invisible artifacts, blocky smears left behind when the encoder cuts corners to save space. Google's new patent describes an AI that scrubs those artifacts out, but without the memory and computing overhead that usually makes AI-based video cleanup impractical.
What Google's AI video cleanup system actually does
Every time a video streams to your phone or TV, the file has been compressed to a fraction of its original size. That compression leaves behind tiny visual flaws: blocky patches, blurry edges, color smearing. A step called "loop filtering" runs after decompression to clean those up before the image hits your screen.
Google's patent describes using a small neural network, a miniature AI, to do that cleanup job. The trick is that different parts of the network are allowed to use different levels of precision. The most important parts get extra numerical detail; the less critical parts run on coarser numbers. That mix lets the AI stay accurate while taking up far less memory and computing power than a standard neural-network approach.
The result is better-looking video frames at a cost that can actually fit inside a phone chip or a streaming device. You would not have to do anything differently; the improvement would happen invisibly inside the video decoder.
producing a filtered block of pixels by applying a heterogeneously quantized depthwise separable convolutional neural network (HQDSCNN) to a decoded block of pixels …
Translation: An AI network cleans up pixel blocks from a video stream.
How the network trims itself to fit tight hardware limits
The patent covers a loop filter built on a type of AI network called a depthwise separable convolutional neural network (a CNN design that does the same image-processing job as a conventional network but with fewer calculations). The specific innovation is how that network is compressed for deployment.
Normally, every number inside a neural network is stored at the same precision, say 8 bits or 16 bits per value. Google's method is heterogeneous quantization, meaning different parts of the network get different bit-widths. Here is how that process works:
- Identify key components: the system finds the layers or filters inside the network that affect output quality the most.
- Set a floor for everything else: non-critical parts get a minimum bit-width, keeping them as small as possible.
- Iterate on the key parts: the system tests candidate bit-widths for the important components, picking whichever combination hits the quality target without exceeding the size or speed budget.
The final network, called an HQDSCNN, applies this tuned, mixed-precision structure to each decoded block of pixels in a video frame, producing a cleaner block as output. Because each layer's precision was chosen specifically rather than set uniformly, the network uses fewer bits overall without sacrificing the quality that matters.
… identifying a set of key components, setting a minimum number of bits for representing non-key components, and iteratively evaluating candidate numbers of bits for representing components of the set of key components …
Translation: The system saves processing power by figuring out which parts of the AI model need the most precision.
What this means for streaming quality on low-power devices
Video decoders run on everything from data-center servers to sub-$50 streaming sticks, and the loop filter runs on every single frame. A filter that is even slightly too heavy can drop frame rates or drain a phone battery noticeably faster. That is why AI-based loop filters have mostly stayed in research: they are too expensive to run in real time on ordinary hardware.
A system that selectively concentrates precision where it counts, and strips it away where it does not, is an engineering approach suited to that real constraint. If this makes it into a future version of a codec like AV1 or its successors, Google's interest in open video codec development could translate into a visible improvement in streaming quality on exactly the low-power devices where quality tends to suffer most, without requiring new hardware.
Google's 59th filing we've tracked since May joins earlier work on editing photos by text and rebuilding video from sample frames in our AI photo editing watchlist.
Blocky edges, smeared faces on video calls, and muddy textures in streamed games are not technical edge cases. They are what billions of people see every day, and they exist because video has to be squeezed aggressively to travel across networks at any reasonable cost.
The cleaning step that repairs those artifacts after compression runs constantly, at enormous scale, and doing it well has historically required more computing power than most devices can afford to spend. Google's approach recognizes that not every calculation inside that process matters equally, so it assigns heavier resources to the parts that drive quality and lighter resources to the parts that don't.
That frugality is the whole argument. A filter that is accurate but too expensive to run everywhere helps almost no one. One that is accurate enough and cheap enough to deploy at the level of widely used video formats touches nearly every screen.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
11 drawing sheets from US 2026/0289247 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in