Does AI upscaling actually work? What it can and cannot recover

The demonstrations are genuinely impressive: a blurry thumbnail becomes a sharp portrait, a soft old photograph gains skin texture and individual hairs. It looks like recovery. It is not.

Understanding what is actually happening is the difference between using these tools well and being embarrassed by them in a context where accuracy mattered.

What upscaling used to be

Traditional upscaling — bicubic, Lanczos, the algorithms built into every image editor — works by interpolation. To invent a new pixel between two existing ones, it takes a weighted average of the neighbours. Lanczos uses a wider, more sophisticated window than bicubic, which is why it looks better, but the principle is the same.

Interpolation cannot add information. If a face in your source occupies forty pixels, no averaging function will produce eyelashes, because there is no eyelash data present — just an average of skin and shadow. Scale up 4× and you get a larger, smoother version of the same forty pixels. Sharpening afterwards increases local contrast, which reads as "crisper" while adding nothing and often introducing halos around edges.

That ceiling is real and it is mathematical. Traditional upscaling makes an image bigger. It cannot make it more detailed.

What AI upscaling does instead

A neural upscaler is trained on millions of pairs: a high-resolution image, and a degraded version of it. Over that corpus it learns the statistical relationship between "what a downscaled eyelash looks like" and "what the eyelash looked like".

At inference it applies that learned mapping. Given forty pixels that look like a downscaled eye, it generates the several hundred pixels that a real eye would most plausibly have produced.

The key word is plausibly. The model is not recovering your eyelashes. It is generating eyelashes consistent with the ones in its training data, conditioned on the low-resolution evidence in your file. When the subject is generic — skin, fabric, foliage, brickwork — the invented detail is statistically excellent and visually convincing, because one patch of realistic skin texture is as good as another.

When the subject is specific, the same mechanism becomes a liability.

Where it works well

Photographs of people, at moderate scale factors. Portraits are the best case. Faces are the most heavily represented category in every training set, and the model has learned skin, hair and eye structure thoroughly. A 2× upscale of a portrait is usually indistinguishable from a genuinely higher-resolution photograph.

Natural texture. Foliage, grass, water, stone, fabric, fur. The eye judges these on whether the texture is plausible, not on whether individual leaves are correct. Nobody can tell that leaf 4,182 is in the wrong place.

Rescuing an image that is otherwise unusable. A 600-pixel product photo that must fill a 1200-pixel slot. The choice is not between an AI upscale and a good original; it is between an AI upscale and a soft, obviously stretched image. The upscale usually wins.

Old photographs with soft focus and film grain. Restoration models are specifically trained on this degradation. They reconstruct faces and rebuild detail that was lost to the original lens, the film stock, and decades of storage. The result is not what the camera saw. It is a plausible reconstruction, and for family photographs that is often exactly what is wanted.

Compressed or noisy sources. Trained on realistic degradation, these models handle JPEG artefacts and sensor noise as part of the same problem, which is why the output frequently looks cleaner as well as larger.

Where it fails, and the failures are specific

Text. The single most reliable failure. A model asked to upscale illegible small text will produce sharp, confident, wrong letters. It has learned what text looks like, not what your text says. The output is highly legible and semantically fabricated, which is far worse than remaining blurry — a blurry document is obviously unreliable, whereas a crisp one invites you to read it.

Never upscale a document, a receipt, a licence plate, a screenshot, or a chart. If you need text out of a low-resolution image, OCR it and keep the extracted text separately, where its uncertainty is visible.

Faces you need to identify. The model reconstructs a plausible face, not that face. It will move the nose slightly, adjust the jaw, and invent an eye colour where there were four pixels. For a family photo this is fine and often lovely. For anything evidentiary, forensic, or identity-related, it is fabrication, and a widely-cited failure mode: upscaling a pixelated face produces a confident portrait of a person who does not exist.

Logos and brand marks. Precise geometry, exact typography, specific proportions. A model will produce something logo-shaped with the wrong letterforms. Always find the vector original.

Fine repeated patterns. Grilles, textiles, halftone print, architectural detail. Models frequently generate moiré or invent regularity that was not there.

Extreme scale factors. At 2× the model is filling in detail with strong evidence. At 8× it is inventing almost everything and the source is a suggestion. Results become smooth, waxy and unmistakably synthetic — skin loses pores, hair becomes strands of plastic.

A rule for deciding

Before running an upscale, ask: would a plausible-but-different detail here cause a problem?

• A jumper's knit pattern, invented differently — no problem.
• A serial number, invented differently — serious problem.
• Leaves on a tree, invented differently — no problem.
• A person's face in a legal exhibit, invented differently — serious problem.

That single question resolves nearly every real case, and it maps cleanly onto the rule of thumb: texture is safe, information is not.

Practical guidance

Start from the best source you have. Upscaling a compressed copy makes the model reconstruct from artefacts as if they were detail. Find the original. It matters more than any setting.

Prefer 2× over 4× where you can. More evidence per generated pixel means fewer inventions. If you need 4×, consider whether 2× followed by a modest conventional resize gets you close enough with fewer hallucinations.

Upscale before other edits, not after. Model output is cleanest from an unmodified source. Sharpening, heavy colour work or aggressive compression beforehand gives it degraded evidence and it will faithfully reconstruct the degradation.

Check the output at 100%, in the four risky places. Any text, however incidental. Faces, especially eyes and teeth. Fine repeated patterns. Boundaries between objects, where models occasionally smear one into another.

Do not upscale twice. Running output back through the model compounds invention on invention, and the second pass has no real evidence at all.

Know which tool you want. These are different jobs with different models:

• Upscaling increases resolution and reconstructs detail.
• Restoration targets old and damaged photographs, with face reconstruction.
• Enhancement fixes exposure, colour, contrast and noise without changing resolution.

Reaching for an upscaler when the actual problem is that the photo is underexposed produces a larger, still-underexposed photo.

The honest summary for each use case

Where this genuinely changes things

The honest case for AI upscaling is not that it defeats physics. It is that a large fraction of images only need to look right at a given size, and for those, plausible detail is functionally as good as real detail.

That covers most photographs most people work with: family archives, product shots, marketing images, social posts. For those, an upscaler is a real capability that did not exist a decade ago.

What it does not cover is any image where someone will act on what the pixels say. There, the model's greatest strength — confident, convincing detail — is precisely the problem, because there is no visual difference between detail it reconstructed correctly and detail it invented. The image gives you no signal about which you are looking at.

Knowing which category you are in before you press the button is the whole skill.

Because these models need GPU-class compute and multi-gigabyte weights, upscaling is one of the few operations that genuinely cannot run in a browser today — OpusImg's AI tools process on a server and say so on each page, in contrast to the compression, resizing and redaction tools that never send your file anywhere. Which tool runs where is worth knowing before you upload something sensitive.