Samsung Patents an Image Quality-Guided Super-Resolution Training Loop
Teaching an AI to upscale images is hard enough — but Samsung's new patent adds a twist: instead of training against a fixed reference photo, the model is guided by a separate AI that scores image quality and picks the best possible target on the fly.
What Samsung's quality-guided upscaling actually does
Imagine you're trying to learn to paint by copying masterworks, but your teacher gets to choose which masterwork you copy based on what will teach you the most at that moment. That's roughly what Samsung is building here for image upscaling.
When a phone or TV upscales a low-resolution photo to a higher-resolution display, it needs to be trained on pairs of blurry and sharp images. The usual approach: pick one sharp reference image and tell the AI "aim for this." Samsung's patent proposes something more deliberate — a second AI, called an image quality assessment model, scores a pool of candidate reference images and picks the best one to train against.
The result is that the upscaling AI doesn't just chase sharpness — it chases perceptual quality, the kind of crispness and detail that actually looks good to a human eye. This kind of training refinement typically happens silently inside a lab and ends up in the camera or display software on your next device.
How the IQA model picks the right training target
The patent describes a two-stage pipeline used during the training phase of a super-resolution (SR) model — not at the moment you take a photo.
First, a low-resolution input image is fed into an image enhancement model (the SR network being trained), which outputs a higher-resolution version. Separately, a pool of ground-truth images — high-resolution reference photos — is run through an image quality assessment (IQA) model. The IQA model scores each candidate on perceptual quality (think: how sharp, natural, and artifact-free the image looks to a human observer, rather than just pixel-level accuracy).
The training loop then:
- Selects the ground-truth image with the best quality score from the pool
- Computes a reference loss — the difference between the SR model's output and that chosen reference
- Backpropagates that loss to update the SR model's weights
The key insight is that not all high-resolution reference images are equally useful as training targets. Some may be noisy, over-sharpened, or compressed. By letting an IQA model curate the reference selection dynamically, Samsung aims to push the SR model toward outputs that score well on perceptual quality metrics rather than just raw pixel similarity (like PSNR or SSIM, which don't always correlate with what looks good to a human).
What this means for Galaxy camera and display upscaling
For Samsung, this is relevant across a wide product surface: Galaxy phone cameras that upscale low-light shots, QLED and Neo QLED TVs with built-in AI upscaling, and even Galaxy tablets rendering streamed content. Better perceptual quality in training directly translates to images that look less "AI-ified" — fewer halos, fewer plastic-skin textures, more natural detail.
The broader implication is a shift from fixed training targets to curated ones. If your upscaling AI learns from only the best-quality reference images — rather than whatever happened to be in the dataset — the resulting model should generalize better to the messy, compressed, real-world images that actually end up on your screen.
This is solid, incremental ML engineering rather than a conceptual leap — but that's not a knock. Training data curation is often where the real gains in perceptual AI quality come from, and building IQA-guided reference selection directly into the training loop is a practical, deployable idea. Samsung ships SR models in a lot of hardware, so even a modest perceptual quality bump has real reach.
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.