Is there an AI tool to increase image quality?
Yes — Real-ESRGAN is free and open source (BSD-3-Clause), with 4x and 2x models plus arbitrary scaling. But understand what it does: AI upscalers don't recover lost detail, they invent plausible detail. Great for a blurry landscape. Legally and forensically worthless for a license plate or a face.
Why — the first-principles explanation
The detail is gone. Not hidden, not compressed, not hiding in a shadow — physically absent from the file. A 200×200 photo contains 40,000 pixels of information, and no algorithm can read the 4 million pixels it doesn't have. This is arithmetic, not a limitation of current tech.
So what does an AI upscaler actually do? It hallucinates, competently. It was trained on millions of pairs — a sharp image and a deliberately ruined copy of it — until it learned the statistics of how real images degrade. Show it a blurry patch and it produces what a sharp version most plausibly looked like, based on every similar patch it has seen. That's why upscaled brick walls look fantastic (bricks are extremely predictable) and upscaled faces look like a stranger who is almost your cousin (faces are high-stakes and every one is different). The model isn't recovering your cousin. It's drawing a face-shaped thing consistent with the smudge.
This is the whole practical rule: the upscaler is an artist, not a microscope. For a landscape, a product shot, a texture, an old scan — the invented detail is plausible and nobody's harmed, and the result genuinely looks better. For anything where the identity of the detail matters — a plate number, a face in a security still, a signature, a serial number — the output is a confident fabrication. Courts and forensic examiners treat it that way, and so should you.
A second thing worth knowing: "quality" isn't one problem. Upscaling adds pixels. Denoising removes sensor grain. Deblurring fixes motion or focus. Artifact removal cleans up JPEG blocking. Tools bundle these, but they're separate operations, and running the wrong one makes things worse — denoising a sharp photo just smears it. Real-ESRGAN targets general image and video restoration: upscaling while reducing degradation from compression, noise, and blur.
On tools: Real-ESRGAN is free, open source under BSD-3-Clause, and runs on your own machine — 4x native, a 2x variant, and arbitrary scales via `--outscale`. Its authors note a real limitation: the portable executable crops the image into tiles and stitches them back, which can leave visible block seams the full PyTorch version avoids. Paid tools (Topaz, Adobe's tooling, phone-maker features) mostly wrap the same class of model in a nicer interface with better tiling.
An example that makes it click
Someone describes a stranger's face to a police sketch artist. The artist draws a face. It's a good drawing — proportional, plausible, convincing.
Now: is that the person? The artist has never seen them. The artist has seen ten thousand faces and drew the most likely one consistent with a vague description. If the sketch looks like your neighbor, that's not evidence about your neighbor. It's evidence about the artist's training.
An AI upscaler is the sketch artist, and your blurry photo is the description. Fine for a poster. Not fine for an arrest.
How to do it
- Name the actual problem first: too few pixels (upscale), grainy (denoise), out of focus or shaky (deblur), or blocky (artifact removal). The wrong tool makes it worse.
- Start with the free option. Real-ESRGAN is BSD-3-Clause licensed and runs locally — no upload, no subscription.
- Use the 4x model for general photos; the 2x variant for gentler enhancement. Use --outscale for in-between sizes.
- Work from the highest-resolution original you have. Never upscale a screenshot of a compressed re-upload — you're amplifying someone else's damage.
- If you see block seams in the output, you're using the portable executable's tiling. Switch to the full PyTorch version.
- Zoom in on faces, text, and numbers before you use the result anywhere. That's where invention shows up first and matters most.
- Never use an upscaled image as evidence about identity — a plate, a face, a signature, a serial number. The detail was invented, not recovered.
- For faces specifically, consider a face-specific restoration model, and still assume the result is a lookalike rather than the person.
Key facts
- Real-ESRGAN is a practical restoration algorithm for general image and video enhancement, released under the BSD-3-Clause license and free for personal and commercial use.
- Real-ESRGAN offers a native 4x model, a 2x variant, and arbitrary scales via the --outscale parameter.
- Real-ESRGAN's authors note the portable executable can introduce block inconsistency because it crops input into tiles, processes them separately, then stitches them together.
- AI super-resolution models are trained on pairs of sharp and degraded images and generate plausible detail; they do not recover information absent from the source file.
- A 200x200 image contains 40,000 pixels of real information — any 4x upscale outputs roughly 640,000 pixels, meaning over 90% of the output is model-generated.
- Upscaled faces, license plates, signatures, and serial numbers are fabrications and are not reliable for identification.
▶ The 60-second explainer (script)
Yes, there are AI tools that increase image quality, and one of the best is free — Real-ESRGAN, open source, runs on your own machine, 4x upscaling. But before you use it, you need to know what it's actually doing, because almost everyone gets this wrong. The detail in your blurry photo is gone. Not hidden. Not compressed. Physically absent. A two-hundred-by-two-hundred image holds forty thousand pixels of information, and nothing can read the four million pixels it doesn't have. That's arithmetic. So what does the AI do? It hallucinates, competently. It was trained on millions of pairs — a sharp image and a deliberately ruined copy — until it learned how real images fall apart. Show it a smudge and it draws what a sharp version most likely looked like. That's why upscaled brick walls look amazing — bricks are predictable. And why upscaled faces look like a stranger who's almost your cousin. It's not recovering your cousin. It's drawing a face-shaped thing that fits the smudge. Think of a police sketch artist. Someone describes a stranger, the artist draws a convincing face. Is that the person? The artist never saw them. They drew the most likely face. If it looks like your neighbor, that says something about the artist's training, not about your neighbor. So: upscaler is an artist, not a microscope. Landscape, texture, old scan — go for it, it genuinely looks better. License plate, face in a security still, a signature — that output is a confident fabrication. Fine for a poster. Not fine for an arrest.
What authoritative sources say
People also ask
What's the best free AI upscaler?
Real-ESRGAN is the standard free option — open source under BSD-3-Clause, runs locally with no upload, and offers 4x and 2x models plus arbitrary scaling.
Can AI recover a blurry license plate?
No. It will produce a sharp, readable, invented plate. The characters come from the model's training, not from your photo, and they are not evidence of anything.
Why does the upscaled face look like a different person?
Because the model is drawing the most statistically likely face consistent with the blur. Fine detail defines identity, and that detail was never in the file to begin with.
Are paid upscalers better than free ones?
Mostly they're friendlier, with better tiling, batch handling, and face-specific models. The underlying approach is the same class of model — the free tools are not far behind.
Why does my upscaled image have visible square seams?
That's tiling. The portable executable crops the input into tiles, processes each separately, then stitches them. The full PyTorch version avoids it.