Enhancing a photo with an AI chatbot: why it changes your picture

What happens when a general AI chatbot “enhances” a photo

You upload a holiday snap to a chatbot and type “enhance this photo” or “make it sharper”. A few seconds later you get something that looks cleaner. Then you look closer: your sister’s smile is slightly different, the sign behind her says something that is almost English, and the file is smaller than the one you started with. This is not a bug. It is how general-purpose AI assistants work when you ask them to edit a picture.

Most AI assistants that chat with images do not edit your file in the way a photo editor does. They look at your picture, build an internal description of it (a woman in a red coat, a harbour, boats, late afternoon light) and then generate a new image that matches that description and the style you asked for. Your photo is used as a strong hint, not as the canvas.

Three consequences follow from that, and they explain almost every complaint people type into search engines:

• The output is small. Image generators usually return around 1-2 megapixels. If your original came from a modern phone, the “enhanced” version can have fewer pixels than what you uploaded. Asking the chatbot to make it bigger produces another new image of roughly the same size, not a true upscale.
• Faces are reinterpreted. The model knows what a face looks like in general, not what your grandfather looked like in particular. It redraws eyes, teeth, wrinkles and hairlines from its training, so the result looks like a plausible person who resembles him.
• Text and small objects drift. Letters on a shop sign, a number plate, the logo on a mug, the pattern on a jumper: anything with precise geometry is regenerated, and generators are still unreliable with exact shapes and lettering.

So when people ask “why does AI change my photo?”, the honest answer is that a generative model never had the job of keeping it. It was asked to produce a nice image in the spirit of yours, and that is what it did. For a social media post that may be fine. For a passport photo, a product listing, an archive scan or a print for the living-room wall, it usually is not.

What a restoration model does differently

A restoration model is a different kind of tool. Instead of describing the scene and drawing it again, it takes the actual pixels of your file as input and predicts, for every small patch, what that patch should look like at a higher resolution or without the noise, blur or damage. The input is the canvas. The geometry of the scene stays where it was.

In practice that means:

• Every original pixel is kept as the reference. Edges, proportions, the position of each eye and each letter come from your photo. The model adds coherent detail between them instead of inventing a new layout.
• You choose the output size. When you ask Airpic to upscale, you get your file at ×2, ×4 or, for large-format print, up to 600 megapixels built tile by tile. A full-resolution phone photo becomes a file a print shop can actually use, not a 1-megapixel redraw.
• The format is respected. JPG, PNG with transparency or 16-bit depth, WebP, TIFF, HEIC and AVIF are accepted, EXIF orientation is applied and the ICC colour profile is preserved, so colours do not shift when you open the result in another program.
• Nothing is added to the scene. A restoration model does not decide that your beach needs a sunset or that a wall would look better without a crack unless you explicitly use a tool for that, such as an object remover with a mask you paint yourself.

Faces are the one area where restoration also has to rebuild detail, because a blurry face simply does not contain eyelashes or iris texture. Here the difference is control. Airpic’s restore old photos tool rebuilds faces with a fidelity slider: lower values keep more of the original face, higher values add more reconstructed detail. You can also add up to three reference photos of the same person so the reconstruction is guided by what they really looked like, rather than by an average face. The tool works on files up to 20 MP.

The same principle runs through the other single-purpose tools. When you remove the background from a photo, the cut-out edge is calculated at full resolution down to single hairs, and the subject itself is not repainted. When you reduce noise in a photo or unblur a picture, an intensity slider lets you stop before the result starts to look artificial.

AI assistant to edit photos or restoration tool: when to use each

None of this means general AI assistants are bad tools. They are excellent at a different job. The useful question is not “which AI is better?” but “do I need a new picture, or do I need my picture, improved?”

Use a generative assistant when:

• You want an image that does not exist yet: an illustration for a blog post, a concept sketch, a mood board.
• You want to change the style on purpose: turn a photo into a watercolour, a comic panel or a poster.
• You are exploring ideas and accuracy does not matter: a background for a slide, a placeholder for a mock-up.
• A small file is enough because the image will only be seen on a phone screen.

Use a restoration tool when the picture has to stay true:

• Photos of real people. Family portraits, old prints of grandparents, wedding photos, a headshot for a CV. A reinterpreted face is a stranger, however pleasant.
• Documents and anything with text. Certificates, receipts, menus, signage, labels, scans of letters. Generators rewrite characters; restoration keeps them.
• Products. If you sell on Amazon, eBay or Etsy, the photo is a description of what the buyer will receive. Changing the stitching on a bag or the print on a T-shirt creates returns and complaints.
• Print. Posters, canvases and banners need real pixels. A 1-2 megapixel image printed on A2 or larger looks soft no matter how good it seemed on screen. Our guide on upscaling an image for large-format printing explains how many pixels you need for each viewing distance.
• Evidence and archives. Insurance claims, property records, historical collections: anything where someone may later ask whether the image was altered.

There is also a privacy angle worth considering when choosing an AI assistant to edit photos. A chat keeps your images in the conversation history unless you delete them, and you need an account to use it. Airpic asks for no account and no email: your file is processed on a GPU in the European Union and deleted at most one hour after you download the result, or after one day if you never do. Images are never used to train models. If you want the full comparison, it is on the why Airpic page.

Why does AI change my photo? How to spot a reinterpreted image

Generated images can be convincing at phone size. The tell-tale signs appear when you zoom in. Open the original and the result side by side, set both to 100 % (one image pixel per screen pixel) and check these areas in order.

1. Teeth. Count them and look at their shape. Reinterpreted smiles often have too many teeth, perfectly even teeth, or a gap that has vanished. A restored smile keeps the original arrangement.
2. Eyes and the direction of the gaze. Catchlights (the small reflections) should be in the same place in both images. If the pupils point in a slightly different direction, the face was redrawn.
3. Hands and fingers. Generators still struggle here: merged fingers, an extra knuckle, a ring that has moved to another finger.
4. Text and logos. Read every word on signs, packaging and clothing. Generated letters look right from a distance and dissolve into near-letters up close. Brand logos may be simplified or replaced.
5. Fine repeating patterns. Brickwork, fabric weaves, window grids, fence posts. A restoration model keeps the count; a generator often adds or removes a row.
6. Background objects. Look for things that appeared or disappeared: a lamp post, a second person, a tree branch.
7. Pixel dimensions. Check the file’s width and height in your file browser. If the “enhanced” file has fewer pixels than the original, it was regenerated, not upscaled.

A quick technique that works well: put the two images in layers, set the top one to half opacity and toggle it on and off. Restored images line up with the original; reinterpreted ones shift, especially around faces and text. Airpic’s result viewer has a before-and-after slider and a “View at 100 %” button, so you can do this check before you download anything.

Upscaling an image with AI without generating a new one: step by step

Here is how to improve a photo without letting the AI change it, using Airpic in a browser on a computer or a phone. No account is needed, and processing and previewing the result are free; you get one free full download a month, and the price for anything beyond that is shown on each tool page.

To make a photo bigger and sharper:

1. Open the tool to upscale an image and drop in your file, or paste it with Ctrl+V. HEIC from an iPhone works as it is.
2. Choose ×2 for a quick improvement for screens, or ×4 for prints and cropping. For large-format printing, switch on the print mode and enter the size you want in centimetres; the tool tells you whether the pixels are enough and suggests a factor up to 600 MP.
3. If the photo has people in it, tick the option to enhance faces.
4. Process, then compare with the slider and the 100 % view. If it looks right, download it in PNG, JPG or WebP.

If the picture is an illustration, a comic or anime artwork with flat colours, use the anime upscaler instead: it keeps clean lines without adding photographic texture.

To restore a face while keeping the person recognisable:

1. Open the tool to restore old photos and upload the scan or photo (up to 20 MP).
2. Leave the fidelity slider in the middle to start with. Very degraded inputs, such as an image forwarded several times through a messaging app, benefit from lower fidelity to avoid a generic face.
3. If you have them, add one to three clear photos of the same person as references. One is enough; more help when the expression or age differs.
4. Optionally switch on recolouring for black-and-white prints.
5. Process and compare the candidates. Check the teeth, eyes and hairline at 100 % using the list above, then download the version that looks most like the person you remember.

For more detail on scans, see our guide to restoring old photos with AI. When you need a new image, a chat is the right tool. When you need your image, sharper and bigger, start by trying to upscale an image on Airpic and compare the result pixel by pixel.

FAQ

Keep reading