IBM Patents Image Compression That Cuts Storage Without Sacrificing Picture Quality

By Patentlyze Team · Updated Jun 26, 2026

IBM's new patent describes a way to compress images by doing the heavy, approximate math cheaply and the fine-tuning math carefully, mixing two very different levels of numeric precision in a single pipeline to get the best of both worlds.

What IBM's mixed-precision image compression actually does

Imagine you're squeezing a huge photo down to a smaller file. Normally the computer does all the crunching at the same level of precision throughout, which is either slow and accurate or fast and sloppy.

IBM's patent describes a smarter split: the first, roughest pass of the math runs in a low-precision format (8-bit floating point, the kind of number-crunching used in AI chips to go fast), while the step that nails down the final quality runs in high-precision mode (64-bit). The idea is that you only need the expensive precision where it actually changes the result.

The end product is a compressed version of the image built from just the most important mathematical "ingredients," called singular values and vectors. You keep only as many of those ingredients as you need for the quality level you want, and throw the rest away. The patent claims this approach reduces both storage space and the bandwidth needed to send images around.

How the sketching and decomposition pipeline splits precision

The method is a variant of Singular Value Decomposition (SVD), a classic technique for compressing images and data by breaking a matrix (a grid of pixel values) down into a ranked list of components, then keeping only the top-k most important ones.

What makes this filing distinct is the mixed-precision sketching step. Rather than running the full SVD directly on the large input matrix, the algorithm first creates a sketching matrix (a smaller, randomized proxy that captures the rough shape of the data) and multiplies the two together using 8-bit floating-point arithmetic. That low-precision multiplication is fast and hardware-friendly, especially on modern AI accelerators.

From that product, a basis matrix Q is extracted using high-precision 64-bit math. That basis pins down the geometry of the image's important content accurately. A smaller matrix B is then derived from Q, and SVD is run on that smaller matrix instead of the full original, yielding:

• Left singular vectors U (how image content is distributed across rows)
• Right singular vectors V (how it's distributed across columns)
• Singular values Σ (how important each component is)

Only the top-k components are kept, effectively discarding the least important detail and producing a compressed representation.

What this means for storage-heavy AI and cloud image pipelines

For everyday photo storage the gains here may be modest, but for large-scale data pipelines, the argument gets interesting. Data centers processing millions of images, AI training systems ingesting visual datasets, or satellite imaging workflows all pay real costs in storage and network bandwidth. A compression method that offloads the bulk of its arithmetic to cheap 8-bit hardware without sacrificing final quality could meaningfully cut those costs.

There's also a hardware angle. The 8-bit precision used in the sketching step matches exactly what modern AI accelerator chips (like those in IBM's own systems) are optimized to run fast. Designing compression algorithms around that reality, rather than treating it as a compromise, is the more interesting idea here.

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