Samsung Patents a Way to Merge Camera Frames Without Ever Fully Decoding Them
Every time your phone takes a burst of photos and merges them into one sharp image, it has to decode each raw frame before it can do any work. Samsung is patenting a way to skip that step and process the frames while they're still compressed.
What Samsung's compressed-frame merging actually does
You're holding your phone up at a dimly lit concert, and it's capturing ten frames in a row to stitch them into one clear shot. The traditional approach decodes every single frame into full-sized pixel data before merging them, which takes memory and processing power.
Samsung's patent describes doing that merging work before the frames are decoded. The phone compresses each raw frame into a compact mathematical summary (called a latent feature set), merges those summaries together in the compressed space, and only then unpacks the final result into a real image. The intermediate frames never have to exist as full-blown pictures at all.
The goal is a faster, more power-efficient pipeline for multi-frame tasks like night mode, HDR merging, and burst photography. Less data moving around means less work for the chip and, in theory, better battery life.
… compressing, using the at least one processing device, the input frames to generate corresponding input latent feature sets in a latent space; performing, using the at least one processing device, multi-frame processing using the input latent feature sets to generate an output latent feature set in the latent space; …
Translation: The device processes compressed data directly without fully decoding the images first.
How the encoder, latent processor, and decoder are arranged
The patent describes a four-step pipeline that sits inside a device's image-processing chain.
- Obtain: The device captures multiple raw frames from the sensor.
- Compress: An encoder (a neural network or similar learned model) converts each raw frame into a latent feature set, a compact representation that retains the image's important information but in far fewer numbers than pixel-by-pixel data.
- Process: Multi-frame operations, such as alignment, noise reduction, or HDR blending, are performed entirely in that compressed latent space, producing a single combined latent feature set. This is the core insight: no decoding is needed to do the merging.
- Decompress: A decoder converts the merged latent set back into a standard pixel-space output frame.
The key technical claim is that the processing happens in latent space (think of it as working on a compact sketch rather than a full-resolution painting). Traditional pipelines decode to pixels first, process, then re-encode if compression is needed again. Samsung's approach keeps data in its compressed form throughout, reducing the bandwidth and memory load on the processor.
The patent is written broadly enough to cover a range of multi-frame tasks, from computational photography to video, though the claims center on still-image frame merging.
What this means for phone camera speed and battery life
For phone cameras, the bottleneck in multi-frame photography is often not the sensor but the pipeline that processes frames afterward. Moving data between memory and the image-signal processor is slow and power-hungry. If Samsung can move the merge step into a compressed domain, the phone does less total work, which could mean faster shutter-to-gallery times and longer battery life under heavy camera use.
Samsung's bet on compressed-domain AI processing would make sense given the company makes both the image sensors and the chips that process them. A tighter integration between the two is the obvious path. For users, the practical payoff would be night-mode and burst shots that feel less like they're making you wait.
Samsung's 124th filing we've tracked since May adds to a run that includes one on a self-testing decoder and one finding duplicate picture regions in our camera sensor push.
Samsung's patent describes combining multiple photos while they are still in their compact, compressed form, rather than unpacking them, processing them, and repacking them. That shortcut, if it works in practice, could mean faster photo processing and less battery drain every time you tap the shutter.
The document describes a goal, not a finished tool. There is no trained software model inside it, no test results, and no indication whether existing phone chips can handle this approach or whether new hardware would be required first.
Camera features like this typically take several years to travel from a filing to something you can actually use, assuming the underlying engineering cooperates. The shortest path to a real product would likely run through software updates to existing hardware, but this document alone cannot confirm that route is 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
19 drawing sheets from US 2026/0278844 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →