Google Patents a Self-Adjusting Compression System That Shrinks Video Files Faster
Every time a video streams to your phone, a tiny probability model is working overtime to guess what comes next in the data. Google's new patent lets that model adjust its own learning speed on the fly, instead of running at a fixed rate the whole time.
How Google's compression tweak affects your video streams
Video compression works by predicting patterns in video data and encoding only the differences, not every pixel from scratch. The math behind those predictions relies on a probability model that tries to guess what comes next. The catch is that most systems update this model at a fixed rate, which is fine on average but wasteful or inaccurate in the moments when the video content changes suddenly.
Google's patent describes a way to dynamically adjust how fast that probability model updates itself, depending on what's happening in the video at any given moment. Instead of a one-speed learning rate baked in at the start, the system calculates a modifier that tweaks the model's update step each time it processes a new piece of data.
For you, this would mostly be invisible. The result is video files that compress more tightly without losing quality, or the same quality at a smaller file size, which matters most for streaming over slow connections or storing large video libraries.
determining an update parameter modifier for a probability model parameter that controls an update function for a probability model for entropy coding the sequence of syntax elements …
Translation: The system calculates a specific adjustment factor to help the video compressor better predict and shrink incoming data.
How the update parameter modifier reshapes the probability model
Entropy coding is the last step in video compression where the encoder takes already-compressed data and squeezes it further using statistics. The core idea is: if the encoder is confident a certain symbol (a chunk of data) is very likely, it can represent that symbol with very few bits. The probability model tracks those likelihoods.
The problem is that probability models need updating as they process new symbols, and the speed of that update, called the update rate or learning rate, is usually fixed. A fast update rate responds quickly to change but is noisy; a slow rate is stable but slow to adapt. Neither extreme is ideal across all types of video content.
Google's approach inserts a step called an update parameter modifier between reading the current model state and applying the update. The modifier adjusts the update parameter (the number that controls how big each learning step is) before the model recalculates probabilities. The patent covers both single-hypothesis models (one probability estimate per symbol) and multi-hypothesis models (several competing estimates that are blended together).
The method works inside existing entropy coding frameworks, meaning it doesn't require redesigning the codec from scratch. It layers on top of the update function that video compression standards like AV1 already use, adjusting the parameter rather than replacing the mechanism.
The probability model is then updated using the update function with the modified probability model parameter.
Translation: The software constantly refines its internal math rules based on the new adjustment to keep the file size as small as possible.
What this means for video codec efficiency at scale
Google controls both the AV1 open video codec and YouTube, which delivers more video per day than almost any other service on earth. Even a small improvement in compression efficiency at that scale translates to real reductions in bandwidth costs and server load. This kind of low-level codec work is exactly where those gains are found, not in splashy new features but in squeezing fractions of a percent more out of every encoded frame.
The patent also signals continued investment in the next generation of video coding tools, likely feeding into whatever follows AV1. For anyone tracking how latest Big Tech patents in video compression are shaping the future of streaming infrastructure, this filing fits a clear pattern of Google refining the statistical engine at the heart of its codec stack.
Google's 628th filing we've tracked since May adds to a run that includes an expressive AI voice and smarter search query sorting.
The problem this patent attacks is real, even if it's invisible to most people. Probability models with fixed update rates are a known limitation in entropy coding, and the mismatch between a static learning rate and dynamic video content does produce measurable inefficiency. That's not a theoretical complaint; it's been documented in codec research for years.
The approach here is narrow and precise, which is appropriate for the size of the problem. This isn't fixing a broken system; it's filing down a specific rough edge. The question is whether the modifier adds enough practical gain to justify the added complexity in an encoder pipeline that already involves dozens of such tuning decisions.
For a company running a video platform at YouTube's scale, even marginal compression improvements justify careful engineering. The filing reads like internal codec work that may already be running in production experiments, not a speculative research idea.
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
10 drawing sheets from US 2026/0254964 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →