Nvidia Patents a Way to Fill in Missing Video Frames Using Motion Data from Past Images
Every time a game or video app renders a frame it didn't fully calculate, there's a gap to fill. Nvidia's latest patent describes a neural-network approach that uses motion data from previous frames to make those filled-in images look convincingly real.
How Nvidia's frame-blending trick actually works
A graphics chip renders a scene. But instead of calculating every pixel from scratch, it looks back at recent frames and uses the motion of objects moving through them to guess what the current frame should look like. That's the core idea here.
Nvidia's patent describes a processor that takes motion information from a stored prior frame, scales it up (a process called upsampling), and combines it with the motion data from the frame being built right now. The result is a new image that reflects how things have moved over time, not just where they are in a single snapshot.
You probably already benefit from earlier versions of this kind of technology if you play PC games: features like DLSS generate high-resolution frames without rendering every pixel at full cost. This patent appears to push that idea further by making the historical motion data more useful in the blend.
upsample motion data from a buffer for a prior image; and cause an image to be generated based, at least in part, on motion data for a current image and the upsampled motion data for the prior image.
Translation: It scales up old movement data to help build a new video frame.
How the processor upsamples past motion data into new frames
The patent centers on a processor with dedicated circuits that do two things in sequence. First, they take motion data stored in a buffer for a previous image and upsample it (meaning they scale it up to match the resolution of the current frame being generated). Second, they combine that scaled-up historical motion data with the motion data for the frame being built right now, and use the combined result to generate the output image.
Motion data here refers to vectors that describe how objects moved between frames: pixel A was here in the last frame and is now there. By stretching that data to cover a higher-resolution grid, the system can make informed decisions about where object edges and textures should land in the new frame.
The patent describes a neural network doing the heavy lifting of the blend. The network takes both data sources as input and produces a reconstructed image that accounts for the trajectory of objects across time, not just their current position.
Key components mentioned in the claim:
- A buffer storing motion data from prior frames
- An upsampling step to match current-frame resolution
- A generation step that fuses current and historical motion data
- One or more neural networks orchestrating the reconstruction
… one or more objects in an image are caused to be generated based, at least in part, on applying one or more offsets to a motion of the one or more objects relative to one or more prior images.
Translation: It creates objects in a frame by shifting their past movements.
What this means for AI-rendered graphics and video quality
The problem this addresses is real and expensive: rendering every pixel of every frame at full quality demands enormous processing power, and it scales badly as resolution and frame rates rise. AI-assisted frame generation has become one of the main ways GPU makers try to close that gap, and getting the temporal part right (how well the system uses information from past frames) is where a lot of the quality difference between good and bad implementations shows up.
For you as a user, better temporal blending means fewer visual artifacts: less flickering around fast-moving objects, sharper edges during camera pans, and smoother video at frame rates your hardware couldn't otherwise hit. the pattern in Nvidia's AI-rendering filings suggests the company is treating temporal accuracy as a core engineering problem, not a secondary polish pass.
Nvidia's 36th filing we've tracked in the GPU rendering race since July builds on quieting multi-chip talk and routing data per chip.
Anyone who has watched a fast-moving object smear or flicker across a screen knows the problem is jarring, and it gets worse as screens become sharper and faster because there is simply more detail to get wrong. Fixing it by recalculating every pixel from scratch is expensive in power and processing time, costs that multiply with every increase in screen resolution.
The approach here uses the history of where objects have been moving to predict where they are going, rather than starting fresh with each new frame. That targets the problem at its source rather than papering over it afterward.
The gap between a patent abstract and a shipping product is real, but the underlying problem shows up in games and video every day, and an approach that scales its ambition to match the actual cost of the problem is pointing in the right direction.
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
51 drawing sheets from US 2026/0301117 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