HISTORY OF UPSCALING

The Evolution of Image Upscaling: From Bilinear to AI (1990–2026)

How did we go from the blurry “Image → Image Size” in Photoshop to AI that adds real detail in seconds? This is the story of upscaling — decade by decade — and why the last five years rewrote the rules forever.

Last updated August 2026

TL;DR — For 30 years “upscaling” meant interpolation: squeezing a colour between two pixels and hoping it looked OK. That gave sharp text in the “before”, but the “after” was always a softer, blurrier, halo‑ridden version of the original. Around 2017, neural networks trained on high‑resolution pairs of the same image changed everything: instead of averaging, they now generate plausible detail. Today that’s Real‑ESRGAN (fast, great for photos and art) and diffusion models like SUPIR (slower, best‑in‑class texture). Read the timeline to see how each step got us here — and why AI finally beats Photoshop.

1990s: Nearest‑Neighbour & Bilinear

The earliest “resize” tools had two honest options. Nearest‑neighbour copied each pixel straight into a block (perfect for pixel art, terrible for photos — the “blocky” look). Bilinear averaged the four nearest pixels, producing the first “soft” upscales: the image got bigger, but the crisp edges turned into a gentle blur. Fast, simple, and the ceiling of what a PC could do in real time.

2000s: Bicubic & Photoshop “Bicubic Smoother”

Adobe added bicubic interpolation — it sampled a 4 ×4 neighbourhood instead of 2 ×2, producing smoother gradients and fewer blocky artefacts. Photoshop’s “Bicubic Smoother” became the default “make it bigger” button and dominated until ~2015. But the fundamental problem never changed: it still only averaged existing pixels, so fine detail stayed soft and the “scaled‑up” look was unmistakable.

2010s: Early ML (Sparse‑Coding & SRCNN)

The first crack at a different approach came from academia. SRCNN (2014) was the first neural‐network upscaler: a small CNN that learned to map low‑res to high‑res on a training set. It beat bicubic, but barely, and was slow. The real leap was sparse‑coding / example‑based methods (like Image Alpha’s “waifu2x”-precursor work): learn a dictionary of patches from a large photo corpus and reconstruct missing detail from similar patches found elsewhere in the image.

These were CPU‑glacial and fiddly to tune, but they proved the core idea: detail can be learned, not just interpolated.

2017–2020: The GAN Revolution (ESRGAN / Real‑ESRGAN)

ESRGAN (2018) introduced generative adversarial networks to upscaling: a generator that upscales, trained against a discriminator that learns to “tell real from fake”. The result looked convincingly sharper than SRCNN. Real‑ESRGAN (2020) fixed ESRGAN’s tendency to invent artefacts by using a Relativistic discriminator and a perceptual loss trained on real‑degradation data (not just synthetic downscales) — so it handles real‑world blurry/compressed photos, not just textbook low‑res.

This is what the site ships: the General model (Real‑ESRGAN for photos, 16 K capable) and the Plus model (the original ESRGAN/RRDB net, tuned for illustrations and anime).

2021–2024: Diffusion & Prompt‑Guided Upscaling (SUPIR, etc.)

GANs are fast but still hallucinate. The next wave used diffusion models — start from noise and iteratively denoise toward a high‑res image, with the low‑res input as a guide. Real‑ESRGAN’s successor wave (and SUPIR, used here as the premium “Diffuser” model) takes a text prompt as an extra guide, so you can steer what kind of detail is added (“a 1970s film photograph” vs “a clean anime line”). Slower and more expensive, but the best texture fidelity — and the first generation where you can direct what the AI reconstructs.

Speed vs Quality: The Generational Trade‑off

Era / Method Typical runtime (today) Quality Notes
Bilinear / Bicubic milliseconds low (blurry) “Good enough” for previews; soft on anything enlarged.
SRCNN / early ML seconds (CPU) medium First learned models; largely superseded.
ESRGAN / Real‑ESRGAN ~1–3 seconds (GPU) high The workhorse: fast, great texture on photos & art. This site’s General/Plus models.
Diffusion (SUPIR) 20–60 seconds (GPU) very high Best fidelity, prompt‑guided, but slow. This site’s premium “Diffuser” model.

Upscale with AI now →

A note on that link: the official tool lives at image‑upscaling.net/upscaling. (Typo‑protection is a good habit — see how much the domain name matters!)

Why AI Finally Beats Photoshop

The decisives were never about raw sharpness — they were about what kind of information fills the new pixels:

The end result: from 2010 to 2026, the default mental model flipped. People stopped accepting “upscaled = blurry” and started expecting AI to actually improve the photo.

Frequently Asked Questions

Is AI upscaling better than Photoshop now?

For enlarging and restoring detail, yes — AI adds texture a resize can’t. For precise, pixel‑perfect compositing or colour work, Photoshop is still the tool. Use AI to get a bigger, sharper base, then Photoshop for the fine‑tuning.

Do I still need to save at high resolution in Photoshop?

Yes — start from the best version you have. AI rebuilds detail that’s hinted at, but it can’t create structure from a featureless blob. A 7 MP photo upscaled 2× beats a 500 ×500 icon upscaled to the same size every time.

Which upscaler should I try first?

Real‑ESRGAN (the site’s General/Plus) for everyday images; a diffusion model (the site’s Diffuser) only when you need top fidelity on a tiny input or want to guide detail with a prompt.

What’s next after diffusion?

Faster samplers, better prompt following, and on‑device models that run in the browser without a server. The trajectory is “higher quality, lower cost, everywhere” — the race that built this whole category.

One‑Line History

Upscaling spent 30 years getting bigger by getting blurrier. Machine learning flipped it: now getting bigger can mean getting sharper, because the model has learned what sharpness looks like. The difference is obvious the moment you compare a bicubic 4× with a Real‑ESRGAN 4× — and it’s why AI upscalers went from a research curiosity to the default “make it bigger” button.

Try the AI upscaler for free →

No account required · Free daily quota · No watermark · Images auto‑deleted

Want the technical architecture? See How AI Upscaling Works (Real‑ESRGAN Explained)]. Practical first steps are in the how‑to guide. Questions? Discord.

© 2024 image-upscaling.net
marvineckhardt04092003@gmail.com
Twitter social icon
Impressum
Marvin Eckhardt
Hinter dem kleinen Dorfe 170
38822 Aspenstedt
Germany
Privacy Policy