Super-resolution
Making a low-resolution image bigger and sharper — by inventing the detail, which is why "enhance" is a lie in every police procedural.
When not to use it
- Anything forensic or evidential. The detail is invented. It looks recovered. Nothing marks the difference.
- Medical imaging, for diagnosis. A generated texture in a scan is a hypothesis rendered as data.
- Identification of people. The model fills faces from its training prior, and that prior is not neutral.
- When you need pixel accuracy. Use interpolation. Blurry and honest beats sharp and invented.
Reach for something else instead
- Bicubic / Lanczos interpolation — invents nothing. The right answer when fabrication is unacceptable.
- Rescanning or reshooting — if the original exists, get the real information.
- Multi-frame super-resolution — combining several real frames adds genuine information rather than inventing it. This is the honest version.
- Accepting the resolution — often fine.
Sources & further reading
- Ledig et al. (2017), Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network — SRGAN; where perceptual quality started beating pixel accuracy, and invention started.
- Blau & Michaeli (2018), The Perception-Distortion Tradeoff — the proof that you cannot have both.
- Wang et al. (2021), Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data — why synthetic degradation breaks on real photos, and how to model it better.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Believing "enhance" recovers information. It fabricates plausible information. The pixels are gone.
- Using GAN or diffusion upscalers where accuracy matters. They're optimised for looking real, which is orthogonal to being right.
- Assuming benchmark performance transfers. Models are trained on bicubic degradation; your photo wasn't degraded that way.
- Not knowing that the perception-distortion tradeoff is a theorem. Sharp and accurate is not a thing you can tune toward.
- Trusting an upscaled face. The model fills from its training distribution, with documented consequences.
At a glance
FieldGenerative AI
What it doesinvents plausible detail, not recovers it
Problem typeill-posed inverse; many originals map to one low-res image
The theoremperception-distortion tradeoff; you cannot have both
Honest optioninterpolation
Neverforensics, diagnosis, identification
DifficultyIntermediate
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Super-resolution vs. interpolation — one invents detail and looks superb; the other invents nothing and looks blurry. Which you want depends entirely on whether someone will rely on the pixels.