Television taught a generation that "enhance" is a button. Zoom into a grainy frame, press it, and a licence plate appears. AI upscalers look like that button finally exists: feed in a soft 512 px render and get back a crisp 2048 px image with pores, fabric weave and individual eyelashes.
The detail is real in the sense that it is on your screen. It is not real in the sense that it was in the original. An upscaler does not recover information. It generates new information that is consistent with what was there, and the researchers who built these tools are clear about the difference.
What Does an AI Upscaler Actually Do?
A classical resize (bilinear, bicubic, Lanczos) only interpolates between the pixels you have. The result is faithful and soft, because a 4x enlargement has sixteen output pixels for every input pixel and fifteen of them have to come from somewhere.
A learned upscaler fills those fifteen from what it has seen in training. The ESRGAN paper (Wang et al., 2018), the ancestor of most tools in use today, describes its own output with unusual candour: it is "capable of generating realistic textures during single image super-resolution. However, the hallucinated details are often accompanied with unpleasant artifacts." Hallucinated is the authors' word, not a critic's.
Its successor, Real-ESRGAN (Wang et al., 2021), targets "blind super-resolution", restoring "low-resolution images with unknown and complex degradations". It is "trained with pure synthetic data": clean images degraded on purpose through "a high-order degradation modeling process" of blur, noise, resizing and compression, so the network learns what clean images tend to look like underneath that damage. That knowledge is what it paints with.
Why Sharper and More Accurate Pull in Opposite Directions
This is not a flaw a better model will fix. It is a proven limit. In The Perception-Distortion Tradeoff (Blau and Michaeli, CVPR 2018), the authors "prove mathematically that distortion and perceptual quality are at odds with each other", and show the result "holds true for any distortion measure."
In practice there are two kinds of "better":
| Goal | What it optimises | What the image looks like |
|---|---|---|
| Low distortion | Closeness to the true pixels on average | Accurate but soft; edges blur where the model is unsure |
| High perceptual quality | Looking like a real, sharp photograph | Crisp and convincing; some texture is a confident guess |
The paper notes that GANs, the method behind ESRGAN and Real-ESRGAN, "provide a principled way to approach the perception-distortion bound." They are built to land on the sharp side of the trade. You cannot have the sharpest possible image and the most faithful possible image at once.
Faces Are Where Invented Detail Matters Most
Face restoration pushes the same idea further. GFPGAN (Wang et al., 2021) "leverages rich and diverse priors encapsulated in a pretrained face GAN", and aims "to achieve a good balance of realness and fidelity." A balance means a trade. The project's own README lists an earlier model's drawbacks as "not very sharp" and "have a slight change on identity", and says v1.4 "produces slightly more details and better identity than V1.3." Better identity is an improvement, and it is also a measured admission that identity can drift.
For a generated portrait that is fine: there is no real person to drift from. For a photograph of a real person, a restored face is the network's best guess at that person.
Courts Have Already Drawn the Line
The distinction has left the lab. In State of Washington v. Puloka (2024), a King County Superior Court judge excluded bystander video that had been processed with an AI enhancement tool. As Quinn Emanuel's review of the ruling summarises it, the enhanced video was found not relevant because it used "opaque methods to represent what the AI model 'thinks' should be shown," rather than showing what actually happened.
That sentence is a good rule for everyone else too. An upscaled image shows what the model thinks should be there.
When Invented Detail Is Exactly What You Want
- Your own renders. A diffusion image has no ground truth. Upscaling a 1024 px render for print adds detail that is as legitimate as the rest of the picture.
- Illustration and anime styles. Flat colour and clean line art upscale well because the prior and the content agree.
- Restoration you will review. Old family photos, where a person who knew the subject checks the face before it is shared.
When It Is the Wrong Tool
- Evidence, insurance claims, inspections or journalism, anywhere the image is a record of what happened.
- Reading things. Text, numbers and plates that are not legible in the source are guesses in the output.
- Identifying people. A restored face can look more like a face and less like the person.
How the Vision Studio Edit Tools Handle It
The guided AI tools in Vision Studio include AI upscale with Real-ESRGAN in a general or an anime model, at 2x or 4x, and face restoration with GFPGAN v1.4. Pick the anime model for line art and flat colour, the general model for everything else, and 2x before 4x: a smaller jump asks the model to invent less.
Edits run on the local engine against the image you are working on, so nothing is uploaded to be enhanced. Whatever tool you use, keep the original file. It is the only version that contains only what was there.
The One-Line Rule
Upscale for beauty, never for proof. If the question is "does this look good?", an AI upscaler is the right tool. If the question is "what was really there?", the answer is in the original pixels, and no model can add to them.
Vision Studio is free, open-source (MIT), and runs locally on your GPU. The upscale, face-restore and other edit tools are documented with their source files at vision-studio-x.com/features. Download it at vision-studio-x.com/download.
-Rocky
#Upscaling #RealESRGAN #GFPGAN #ImageEditing #LocalAI #OpenSource #VisionStudio