Samsung Patents a Neural-Network Image Codec Built Around Parallel Processing Layers

By Patentlyze Team · Updated Jun 5, 2026

Samsung is filing patents on a fully learned image codec — meaning a neural network, not a hand-crafted algorithm like JPEG or HEIC, does all the compressing and decompressing. The twist is an architecture called WaveResNeXt that runs its processing layers in parallel rather than sequentially.

What Samsung's AI-driven image compression actually does

Imagine your phone's camera saving a photo at half the file size of a JPEG but with noticeably fewer of those blocky artifacts around edges and fine detail. That's the promise of "learned" image compression — where a trained neural network figures out how to shrink your photo instead of following the fixed rules baked into older formats like JPEG or WebP.

Samsung's patent describes a specific neural-network design, called WaveResNeXt, meant to do exactly that. It takes your image, crunches it down to a compact internal representation, and then uses a second network to model the statistical patterns in that representation — making the final compressed file even smaller without losing more quality.

The decoder side runs the whole process in reverse, reconstructing a clean image from the compressed bitstream. The parallel-processing design is the key architectural bet here: instead of doing each step one after another, several operations run at the same time, which can make encoding and decoding faster on modern hardware.

How WaveResNeXt encodes and reconstructs an image

The patent describes an encoder–decoder pipeline built around the WaveResNeXt architecture. The encoder maps an input image to a latent representation (a compact, learned feature space — think of it as the network's internal shorthand for the image) using layers that run in parallel rather than in sequence.

A separate hyper encoder then takes that latent representation and produces a hyperprior — a statistical model of the latent data (essentially: what patterns does this compressed version tend to have?). This is a well-established trick in learned codecs; knowing the statistics of your data lets you squeeze even more redundancy out during entropy coding (the final lossless compression step that turns everything into a bitstream).

On decode, the process reverses:

• The bitstream is entropy-decoded using the hyperprior as a guide.
• The decoder reconstructs the image from the quantized latent and hyper-latent representations.
• The result is a reconstructed image that should be visually close to the original at a significantly lower bit rate than legacy codecs.

The "WaveResNeXt" name suggests the architecture draws on wavelet-style feature decomposition and ResNeXt-style grouped convolutions (a way of splitting computation across parallel paths, originally from image classification research). That parallel structure is likely the core efficiency claim.

What this means for next-gen image quality and file sizes

Learned image codecs have been closing in on — and in some benchmarks beating — established standards like HEIC and AVIF for a few years now. The bottleneck has been compute: neural compression is expensive to run. Samsung's parallel-layer design is a direct attempt to close that gap, making the codec practical enough to run on device rather than just in a research lab.

For you, this matters because Samsung ships hundreds of millions of Galaxy phones and televisions. If a learned codec like WaveResNeXt matures into a real product feature, it could mean your photos and video frames take up meaningfully less storage and bandwidth — without the visible degradation that comes from cranking up compression on JPEG or even HEIF.

Editorial commentary on a publicly published patent application. Not legal advice.