Samsung Patents an AI Method for Creating Slow-Motion Video From Normal Footage
Slow-motion video usually requires a camera that physically captures more frames per second. Samsung's new patent takes a different approach: instead of recording extra frames, it invents them using AI, doing the heavy math in a compressed form of the video rather than on raw pixel data.
What Samsung's slow-motion trick actually does to your video
Every time you record a slow-motion clip on your phone, the camera has to capture many more frames per second than a regular video. Most phones can do this for a few seconds, at lower resolution, before the processor overheats or storage fills up. Getting true high-resolution slow motion is genuinely hard.
Samsung's patent describes a system that sidesteps the hardware limit entirely. Instead of requiring the camera to shoot faster, the software takes two ordinary consecutive video frames and synthesizes the frames that would have existed between them. The result looks like slow motion even though your camera only recorded at a normal speed.
The clever part is where this synthesis happens. Rather than working on the full, heavy pixel image, the system first compresses the frames into a much smaller mathematical representation, does its frame-inventing work there, and then expands the result back into a real image. That compression step is what makes the process light enough to run on a phone.
… compressing, using the at least one processing device, two adjacent input video frames to obtain corresponding input latent feature sets in a latent space; generating, using the at least one processing device, intermediate latent feature sets in the latent space based on the input latent feature sets; …
Translation: The system shrinks two regular video frames into compressed data models and creates missing middle versions between them.
How Samsung generates frames in compressed space, not pixel space
The patent describes a four-step pipeline for generating slow-motion video from standard footage:
- Compress to latent space: Two neighboring video frames are encoded into a compact mathematical representation called a latent feature set. Think of it as squashing the image down to its most essential information, the way a ZIP file works but tuned for video content.
- Generate intermediate latents: The system then creates new latent feature sets that sit logically between the two compressed inputs. This is where an AI model estimates what a frame at that moment in time would look like, working entirely in the compact compressed form.
- Decompress to pixel frames: Those invented latent sets are then decoded back into full pixel images, called interpolated frames.
- Insert into the timeline: The interpolated frames are placed between the original two frames, producing a video that plays back at a lower apparent speed without any camera hardware change.
Doing the AI inference in compressed latent space rather than on raw pixel data is the central engineering bet here. Pixel-space interpolation (the older approach) requires processing every pixel individually, which is expensive. Latent-space processing reduces the dimensionality of the problem significantly, lowering compute cost and making on-device processing more practical on a mobile chip.
What this means for slow-motion video on everyday phones
For Galaxy phone users, a mature version of this could mean high-resolution slow motion in everyday video mode, not just a special low-res setting you have to enable before something happens. You could theoretically apply the effect after the fact, to footage you already shot.
The practical ceiling on quality is real, though. AI-generated frames are educated guesses, and motion that is fast, unpredictable, or complex (a soccer kick, splashing water, a pet darting across a room) gives the model less to work from. the pattern in Samsung's video AI filings suggests this is one piece of a larger bet on computational video quality, but the slow-motion use case is where the gap between hardware limits and user expectations is most obvious.
Samsung's 133rd filing we've tracked since May adds to a run of ideas, including frame-by-frame movement tracking and past scenes fixing blurry footage, in our camera sensor push watchlist.
The core cost of this approach is accuracy. Instead of recording what actually happened, the software reconstructs what probably happened between two frames, which works beautifully for smooth, predictable motion and falls apart for anything chaotic or fast-changing.
That tradeoff reads as worth making for most people. Birthday candles, a child's first steps, a game-winning moment, these do not need forensic precision, just something that looks convincing on a phone screen.
The real question the patent leaves unanswered is whether working inside compressed data actually delivers a quality jump big enough to replace dedicated slow-motion camera hardware. The architecture is clever, but without published quality results, only a real demo can settle that.
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
14 drawing sheets from US 2026/0292098 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