Sony Patents a System That Blends Two AI Models to Restore Compressed Images
Every time a photo or video gets compressed, it loses something. Sony's new patent describes a way to let two AI models compete to find the best possible fix, automatically.
How Sony's AI repair system handles compression damage
Imagine you take a high-quality photo and send it over the internet. By the time it arrives, it's been compressed, and that compression leaves behind blurring, blockiness, and other artifacts. The photo looks worse than what you originally captured.
Sony's patent describes a system where two separate AI models, each trained to restore image quality in different ways, work together to clean up that damaged image. Instead of picking one model or the other, the system tries blending them at different ratios and then uses a third AI judge to decide which blend looks closest to the original.
The result is a set of instructions that tells any decoder exactly how much of each AI model to apply when restoring your image. Sony's approach is designed to do this at a lower processing cost than running a single heavyweight model from scratch.
How the two AI models compete to find the best blend
The patent covers both the encoding side (compressing an image) and the decoding side (reconstructing it). The core idea is a mixing system built on top of a generative adversarial network (GAN), a well-established AI training setup where a "generator" tries to create convincing images and a "discriminator" acts as a critic judging whether those images look real.
Here's the sequence:
- An original image is compressed into encoded data using standard techniques.
- That encoded data is immediately decoded back, producing a lower-quality "decoded image" that shows compression artifacts.
- Two pre-trained AI image-processing models are applied to that decoded image at two different blend ratios, producing two candidate restored images.
- The GAN's discriminator evaluates both candidates against the original and decides which blend ratio produces the most faithful result.
- That winning ratio is packaged into "control instruction data" and sent alongside the compressed image file.
When a viewer's device decodes the image later, it reads those instructions and applies the exact same AI blend to restore quality. The two models stay lightweight because the heavy optimization work happened once at encoding time.
What this means for video streaming and image quality
Video compression is everywhere: streaming services, video calls, social media uploads, and broadcast television all rely on it. The artifacts it introduces (blurry edges, color banding, blocky shadows) are a constant quality ceiling. Sony's approach treats the encoder itself as the quality-control stage, doing the expensive AI optimization once so the decoder doesn't have to.
For consumer electronics, this fits squarely into Sony's camera, television, and PlayStation ecosystems. A camera could embed the optimal AI restoration instructions at capture time, and a TV or phone could apply them cheaply on playback. It's a practical approach to squeezing better image quality out of the same compression standards without waiting for a new codec.
This is solid, incremental engineering work rather than a dramatic leap. Sony is essentially using GANs to pre-compute the best AI restoration recipe at encoding time and ship it with the file, which is a genuinely practical idea for their hardware ecosystem. It won't rewrite how codecs work, but it could improve image quality on Sony cameras and Bravia TVs without demanding more from the viewer's device.
The drawings
24 drawing sheets from US 2026/0220823 A1 · click any drawing to enlarge
Which company should we read for you?
We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.
Get one Big Tech patent every Sunday
Plain English, intelligent commentary, no hype. Free.
Editorial commentary on a publicly published patent application. Not legal advice.