Get the newsletter
Samsung · AI / ML · Filed Nov 25, 2025 · USPTO published Jun 4, 2026 · Jun 4 release

Samsung Patents an AI Image Compression Method That Fights Quantization Errors

Every time an image is compressed, tiny rounding errors chip away at quality — and most codecs just ignore them. Samsung's new patent describes a learned compression system that bakes awareness of those errors directly into the encoding process.

By · Updated · 3 min read

Samsung Patent: Quantization-Aware Learned Image Compression, figure from US 2026/0156266 A1
Figure from the official USPTO publication.
Publication number
US 2026/0156266 A1(US20260156266A1)
Applicant
Samsung Electronics Co., Ltd.
Filed · Published
Nov 25, 2025 · Jun 4, 2026
Inventors
Saifeng Ni, Madhukar Budagavi, Indranil Sinharoy, Rajan Laxman Joshi
Status when we published
Waiting for an examiner
Document
20 claims
US classification
375/240.02
Parent application
Claims priority from a provisional application 63727120 (filed Dec 2, 2024)
AI training & infrastructure

What Samsung's quantization-aware image compression actually does

Imagine you're saving a photo to send over text. Your phone has to squeeze that image down into a smaller file, and every time it does, it makes little rounding decisions — like rounding 3.7 down to 3 — to fit the data into fewer bits. Those small rounding mistakes, called quantization errors, add up and quietly degrade your image.

Samsung's patent describes an AI-based compression system that doesn't just ignore those rounding errors — it accounts for them as part of the compression process. Before the rounding happens, the system applies a customized "gain" (think of it like a tuning knob) to each data channel in the image's compressed representation, so the errors that do occur matter as little as possible to the final result.

The end goal is a reconstructed image that looks closer to the original, even after the whole compress-transmit-decompress cycle. It's essentially teaching the codec to be a more careful packer.

How the gain-channel system reduces compression artifacts

The patent describes a learned image compression pipeline — meaning a neural network handles the encoding and decoding rather than a traditional algorithm like JPEG or HEIC.

Here's how the pipeline works:

  • An input image is passed through an encoder network that maps it to a latent representation (a compact, abstract description of the image's content, generated by a neural net).
  • Before quantization, a learned gain is applied per channel — each channel in the latent space gets its own scaling factor, tuned to minimize the damage that rounding will cause.
  • The latent is then quantized (rounded to discrete integer values) to make it compressible, and encoded into a bitstream using entropy encoding (a lossless compression step that exploits statistical patterns in the data).
  • On the decoding side, the bitstream is reversed back into a reconstructed image.

The key insight is that by learning per-channel gain factors, the system can scale each channel's values into a range where quantization rounding does the least perceptual damage. It's a smarter front-end to the quantization step rather than a fix applied after the fact.

What this means for AI-compressed photos and video

Learned image compression is an active research area, and the gap between neural codecs and traditional ones (JPEG, WebP, HEIC) is closing fast — especially at low bitrates. A technique that explicitly reduces quantization error could push reconstructed image quality noticeably higher without requiring more bits, which matters a lot for bandwidth-constrained scenarios like mobile photo transfer, streaming thumbnails, or satellite imagery.

For Samsung specifically, this is relevant to its device ecosystem — Galaxy phones, cameras, and display hardware all touch image compression pipelines. If this approach migrates into on-device encoding, your photos could compress more cleanly before ever leaving your phone.

Source: US 2026/0156266 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Samsung patent news breaks.

Add Patentlyze on Google One click.
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.

More from Samsung

All 1,629 Samsung patents →

More in AI training & infrastructure

All 316 AI training & infrastructure patents →