Can you use AI to upscale images?

Short answer: when you ask ChatGPT to upscale a photo, it makes a fresh image based on yours. That works for a sharper-looking holiday snap. It is the wrong tool when every detail has to match the original.

What ChatGPT actually does when you ask it to upscale

OpenAI describes ChatGPT Images as a way to create new images and edit existing ones: you upload a picture and describe the changes you want. Its help page covers editing, a selection tool, an aspect ratio picker and templates. It does not describe an upscaling or “increase resolution” feature. It also warns that highlights are not always precise and edits may extend beyond the area you selected, and it says that changing the aspect ratio regenerates the image.

So a request to “make this higher resolution” is handled as an edit. The model uses your photo as a reference and produces a new image that resembles it. Often close, never guaranteed identical. OpenAI treats the result as a generated image too: images made with ChatGPT carry Content Credentials (C2PA metadata) and a SynthID watermark.

This applies to every method, ChatGPT included. A low-resolution file holds only so much information. A traditional upscaler resamples the pixels you have, which makes the image bigger and usually softer. An AI upscaler uses a model to add plausible detail. Any detail that was missing from the original is a guess.

What you get back: size, sharpness and what changes

OpenAI’s help page says you can generate images in any aspect ratio, using the picker or your prompt, and its API documentation says output size, quality, format and compression can be adjusted. Neither page gives pixel dimensions for ChatGPT, so we don’t either. Treat any fixed figure on a tool-comparison page with caution: download the result and check its dimensions yourself.

Sharpness usually goes up, because the model draws crisp edges where the original had blurry ones. That is the appeal and also the risk, because a sharp detail can still be wrong. Before you trust a result, put it side by side with the original, zoom in and check:

• Faces. Eyes, teeth, nose shape, hairline. Small drifts can make someone look like a relative of themselves.
• Text. Signs, labels, slogans, book spines. Generated lettering can look sharp and say something else.
• Logos. A redrawn brand mark is no longer the brand mark.
• Numbers. Dates on a receipt, digits on a clock, a house number or number plate. Check them character by character.

If any of these matter and any of them changed, you no longer have a faithful copy.

How to do it well

1. Start from the best original you have. For a print or an old photo, rescanning at a higher resolution captures real detail that no upscaler has to invent.
2. Keep the untouched original and work on a copy.
3. Ask for restraint. “Reduce the noise and sharpen this photo. Keep faces, text, logos and numbers exactly as they are” works better than “make this 4K”. It will not always comply.
4. Run the four checks above. If something changed, say what and ask again, or use the selection tool to limit the edit to one area. OpenAI notes the edit may still spill over.
5. Think before you upload. ChatGPT’s data controls let you choose whether your conversations help train OpenAI’s models: turn off “Improve the model for everyone” under Settings, Data controls. You can also use a temporary chat, which OpenAI says is not used to improve its models but may be kept for up to 30 days for safety. Don’t upload other people’s private photos at all.

For wider editing and restoration jobs, see using AI to edit photos.

When to use a dedicated upscaler instead

A dedicated upscaler is built to enlarge the image you give it, and usually lets you set how far to enlarge and how much detail to add. Choose one when fidelity matters: a product shot, an archive scan, a document, or anything you plan to print large. If the tool uploads your photo to its own servers, the same privacy questions apply as with ChatGPT, so check before you use it.

ChatGPT suits the casual end: a holiday snap for a group chat, a social post, a mood board, or a time when you want a restyled version anyway. Whichever you choose, run the same checks, because an AI upscaler that adds detail is guessing too. No tool recovers detail the camera never captured.

Rules and risks

Passport and ID photos. GOV.UK5 says a digital passport photo must be “unaltered by computer software”. A photo upscaled or sharpened with AI breaks that rule. More in using AI for a passport photo.

Evidence. If a picture is meant to show what happened or who someone is, for an insurance claim, a dispute or a complaint, an AI-regenerated version is a different picture. Hand over the original, and if you also include an enhanced copy, label it as one.

Other people’s images. The IPO says photographs and other images are generally protected by copyright, and you will usually need the owner’s permission to copy them or share them online. Upscaling someone’s photo makes a copy of it, and making it bigger or sharper gives you no rights in it. The IPO also says deliberately removing metadata that identifies the copyright owner is unlawful. If you’re tempted to tidy away a watermark as well, read removing watermarks with AI first, and is AI art illegal for the wider legal picture.

Disclosure. Plain upscaling usually needs no label. YouTube, for example, lists video sharpening, upscaling or repair among changes that do not need disclosing, while realistic changes to real people or events do. If an “upscale” altered a face or a detail someone would rely on, say so. ChatGPT’s provenance signals mark its images as generated, but OpenAI notes Content Credentials can be lost through a screenshot or file conversion, and says the signals do not guarantee that content is accurate or unedited.

This is general information, not legal advice.