Faithful vs generative upscaling: which one changes your photo

A small or damaged photo is missing information. Some of it can be inferred with high confidence: a blurred edge is still an edge, compression blocks follow a known pattern, and noise is random enough to separate from the picture. Some of it cannot: the exact strands of hair in a face 30 pixels wide, the letters in a word that became a smudge, the weave of a fabric that compression flattened.

A faithful tool, also called a restoration model, reconstructs the first kind and leaves the second alone. Where information is there, it sharpens; where it is not, it stays smooth. A generative tool fills in the second kind as well. It has learned what hair, letters and fabric usually look like and draws a plausible version wherever the photo is unclear.

Most real tools mix both, and some let you choose how creative they should be. The useful question is not which technology a tool uses, but how much of what you see in the result came from your photo.

Restoration models are usually trained on pairs: a clean photo and a damaged copy of it. The model learns to turn the copy back into the clean photo and is penalised whenever its output differs from the real one. That pushes it toward answers the input supports. When it is unsure, the safest answer is a smooth one, which is why faithful results can look soft or flat where information is missing.

Generative models are trained to produce images that look real, often guided by the input image or by a text prompt. They are rewarded for convincing detail, not for matching an exact original. That is how they turn a tiny, blurry face into a crisp portrait, and also why the portrait may not be the same person.

Identity changes. People notice the smallest differences in faces. A generative enhancer may change the shape of the eyes, the line of the nose, the teeth or the skin. The result looks like a sharp photo of someone similar. For family photos and profile pictures, that is a real loss.

Invented text. When letters are unclear, a generative model draws letters that fit the shapes. A street sign, a receipt, a licence plate or a document can end up with characters that look certain and are wrong. Because the result looks sharp, the error is easy to miss.

Products that no longer match. On a listing, generated detail can add a knit pattern, a texture or a label line the real item does not have. Shoppers buy what they see.

Records and evidence. Photos for insurance claims, property condition, science, journalism or legal purposes must show what was there. Any generated detail, even subtle texture, undermines that.

Generative tools are not bad tools; they solve a different problem. For artwork, concept images or decorative prints, plausibility is the goal. If you upscale AI-generated art, there is no real original to stay faithful to.

A milder case is texture. A night photo whose phone smeared the grass into flat green looks wrong to the eye even when it is 'accurate'. Adding natural grass-like texture there makes it look like a photo again. That is a reasonable trade for a keepsake, as long as you know the texture is plausible rather than recovered.

Use a tool that adds nothing whenever exact detail matters more than looks: documents, receipts, screenshots, licence plates, serial numbers, evidence, and anything you will compare with the real thing.

The trade-off is that strictly faithful results are limited by the input. A very small or damaged photo comes out cleaner and larger, but flatter than a generative result, because the tool will not draw what it cannot see.

A few simple checks show how much a tool adds:

• Shrink the result back to the original size and compare. A faithful result should look almost identical to the input. Features the input does not have were added.
• Use an image with small text you know. Crisp but wrong words mean the tool is generating.
• Use a photo of someone you know well, with the face small in the frame. Ask whether the result still looks like them, not just whether it looks sharp.
• Look at areas that were flat or blurry in the input. New texture appearing there is generated.
• Run the same image twice. Tools that sample randomly give different fine detail on each run.

We sit close to the faithful end, but not at it, and we would rather say so plainly.

Our three main models, Enhance, Upscale 4× and Upscale 8×, work in two stages. A restoration network removes noise and compression and rebuilds edges that the pixels support. A detail stage then adds natural fine texture, such as fabric weave, foliage or the grain of wood, where the photo had lost it. That texture is plausible, not true. The models are tuned not to restyle a photo or reshape faces, but you should still check faces at 100 % zoom.

In Enhance the detail stage is gated. On a photo that is already clean, it leaves the picture essentially as it is. It detects text and leaves it to the restoration step, which removes noise and compression and does not redraw letters. Detection can occasionally miss very small or dense print, so for documents where every character matters, check the result or keep the original. With Upscale 4× and 8×, small text stays as soft as in the original; we don't guess letters.

If you want nothing added at all, for example on screenshots, graphics or text where every pixel counts, keep the original or make a fresh capture: every one of our models can add some fine texture.

None of our tools are suitable for forensic or evidential use, and none will make unreadable text readable or reconstruct a 50-pixel face.

You get 3 free credits a day, with no sign-up and no card. One photo with one tool uses 1 credit; Upscale 8× uses 2. Free uploads can be JPG, PNG or WebP up to 3 MB and 8 MP, and free results carry a watermark. Pro ($9.99 a month, 200 credits) and Pro+ ($19.99 a month, 500 credits) remove the watermark and accept uploads up to 20 MB and 30 MP. Uploads and results are deleted within 24 hours and are never used to train AI.