Google Patents Technology That Reconstructs Video by Layering Motion and Still-Image Estimates
Every time you watch a compressed video, your device is guessing what most of the picture looks like rather than storing every pixel. Google's latest patent describes a way to make those guesses significantly more accurate by combining two approaches that have historically been used separately.
How Google wants to fill in compressed video frames
Video compression works by storing only the differences between frames rather than every frame in full. To do that, a video decoder has to predict what each small block of the picture looks like before it even receives the real data. Today, decoders typically use one prediction method at a time for each block, which means they sometimes pick the wrong tool for the job.
Google's patent describes a decoding method that runs two prediction types simultaneously for the same block, then blends them together. One type looks at how nearby blocks within the same frame are shaded or textured (called intra prediction). The other looks at how a similar block moved between earlier frames, accounting for distortion or warping in the motion (called warp inter prediction). Combining both gives the decoder a richer starting estimate.
The result is that the final corrective data the encoder has to transmit is smaller, because the prediction was already closer to the truth. For you as a viewer, that could mean sharper video at the same file size, or the same quality at a lower data rate.
accessing, from an encoded bitstream, compound warp inter-intra mode data indicating that the current block is coded using compound warp inter-intra mode; accessing, from the encoded bitstream, intra prediction mode data indicating an intra prediction mode for the current block …
Translation: The system reads specific instructions from the video file to determine how to blend motion and still image data.
How the compound warp inter-intra decoder combines signals
The patent covers a decoding process called compound warp inter-intra prediction. When a video decoder encounters a block tagged with this mode in the bitstream, it runs two separate prediction routines and blends their outputs before applying the residual (the small error-correction signal that makes the final image accurate).
The first routine is intra prediction, which estimates a block's contents by looking at the pixels already decoded in the same frame, typically the edges or gradients just outside the block. The second is warp inter prediction, which finds a matching region in a previously decoded reference frame and applies a geometric warp transform (a mathematical adjustment for perspective shifts, zoom, or distortion) to map that region onto the current block's position.
- The decoder reads a special flag from the bitstream confirming compound warp inter-intra mode is active.
- It also reads an intra prediction mode value that tells it which direction or pattern to use for the intra estimate.
- Both prediction outputs are blended (weighted and summed) into a single predicted block.
- The encoder's stored residual data is decoded and added on top to produce the final reconstructed block.
The key engineering claim is that combining these two independently derived estimates produces a more accurate starting prediction than either could alone, leaving less error for the residual signal to correct and potentially shrinking the compressed file.
… obtaining predicted block data for the current block by combining the warp inter prediction block data and the intra prediction block data, obtaining decoded block data by decoding encoded block data accessed from the encoded bitstream, and including, in the reconstructed block data, a sum of the decoded block data and the predicted block data.
Translation: The software creates a final image by adding together the predicted motion, the still details, and the raw video data.
What this means for streaming video quality
For streaming services and video call apps, compression efficiency translates directly into picture quality at a given bitrate. A decoder that predicts frame content more accurately needs fewer bits per frame to hit the same visual quality, which matters most in low-bandwidth environments like mobile networks or congested Wi-Fi. You would notice the payoff in scenes with complex motion or mixed textures, where current codecs tend to produce the most visible artifacts.
This patent fits squarely into the ongoing competition around open video codecs. Google has been a leading contributor to the AV1 codec, which the patent's technical classifications suggest this work may extend or inform. The practical question is whether the added decoding complexity is cheap enough to run on everyday hardware in real time. Codec improvements like this one sit alongside a steady flow of new Big Tech patents pushing video compression forward, as companies race to deliver higher-resolution streams without ballooning data costs.
Google's 43rd filing we've tracked since May in the AI photo editing race builds on erasing people from shots and a noise-removal system for AI video.
The blocky, smeared distortion that spreads across someone's face during a weak video call happens because the decoder is working with a single, failing prediction about what the image should look like. This patent describes a method where two complementary predictions run at once per image block and get blended together, giving the decoder more to hold onto when conditions get rough.
For the person on the call or watching the stream, this means video holds together longer before it visibly falls apart, particularly during the moments that matter most: a bad signal, a crowded network, a stream fighting for bandwidth.
The honest uncertainty is cost. Running two predictions instead of one takes more processing power and battery, and whether the phones and budget devices most people actually use can absorb that load without adding frustrating delay is a real question this document leaves open.
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 2026/0254943 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →