AI Photo Editing for Product Images: Background Removal, Object Removal and Upscaling

AI photo editing is the umbrella term for work done by models that make predictions about the pixels in a photo: removing a background, erasing an unwanted object, upscaling a low-resolution image, colorizing a black-and-white photo and restoring a damaged one. In every one of these tasks the model fills in information the photo does not contain. When it removes a background, it decides where the product's edge ends; when it erases an object, what was behind it; when it upscales, what the missing detail looks like. Cropping, resizing, format conversion and compression need no prediction at all: classical image processing handles them and returns the same result for the same input every time.

This guide draws two lines for businesses that prepare product images: which work goes to AI and which goes to rule-based processing, and which images AI should not touch at all. The second line fits in one sentence: the product's color, shape and label stay exactly as photographed; AI edits what surrounds the product, and the result of any step that touches the product itself, such as upscaling, is compared with the original.

What AI photo editing actually does

There are five core tasks. Each runs on a different kind of model, and each fails in its own way:

• Background removal: The model performs segmentation, an image processing task that predicts whether each pixel belongs to the product or the background, and returns the result as a mask. Everything outside the mask becomes transparent or is filled with a color you choose. Apple's Vision framework does the same job with its foreground instance mask request: it returns a mask that separates the prominent objects in an image from their background.
• Object removal: You mark the area to erase, and the model fills it with new, plausible pixels that match the surroundings. The technical name is inpainting. The filled area is a prediction: the model does not know what was really behind the object; it generates what could be there. Adobe's walkthrough of Generative Fill in Photoshop shows exactly this: the selected object is replaced with new AI-generated pixels, and every run offers three different variations.
• Upscaling and sharpening: Super-resolution models increase the pixel count and fill in the missing detail with patterns learned from similar images. The result looks sharp, but the added detail did not come from the photo; details with an exact shape, such as text, barcodes, logos and faces, can distort. The paper behind Real-ESRGAN, a widely used open-source upscaler, lists among its limitations twisted lines in some results, unpleasant artifacts introduced by GAN training, and complex real-world degradations it cannot remove and may even amplify.
• Colorization: The model assigns a likely color to each region of a black-and-white photo. It guesses typical colors such as sky, leaves and skin convincingly; it cannot know the real color of a fabric, a car or a shop sign.
• Photo and face restoration: The model removes scratches, stains, tears and blur; face restoration redraws a blurred face. The redrawn face resembles the person, but the eyes, mouth and facial features can carry details the person does not have.

The job of classical image processing: crop, size, format and compression

Cropping an image to 1:1, resizing it to 2000 pixels wide, converting it from PNG to WebP or shrinking the file needs no AI model. These operations follow fixed rules: the same image with the same settings gives the same result every time, and the result is known in advance. That predictability is what makes repeated work reliable, such as bringing hundreds of product images to the same size. Using AI where it is needed, and not where it is not, lowers both cost and error.

The thing to watch when choosing a format is transparency. A cut-out image stays transparent only if it is saved in a format with an alpha channel, PNG or WebP; JPEG has no alpha channel, so the transparent area is filled with a flat color (MDN image format guide). According to Google's WebP documentation, WebP supports transparency in both lossy and lossless compression and often produces a noticeably smaller file than PNG for the same image. If a sales channel wants a white background, place the image on white deliberately and save it that way, rather than leaving it transparent and letting the channel fill it with its own color.

Background removal for product images

Background removal is the most common and least risky use of AI on product images, because it does not touch the product; it only decides where the product ends. All the risk sits on that edge, and the hard cases are well known:

• Fine, irregular edges: hair, fur, feathers, tassels and knitted textures. The mask either cuts into them or pulls part of the background into the product.
• Transparent and translucent products: glass bottles, plastic packaging, tulle. Whether the background seen through the product belongs to the product is a question the model cannot settle on its own.
• Shiny and reflective surfaces: metal, glazed ceramics, screens. The surroundings reflected on the surface can be removed with the background, or stay inside the product.
• A product the same color as its background: white on white, black on dark. The lower the contrast, the harder the edge is to predict.
• Shadows: The shadow is often removed with the background, leaving the product floating. If the shadow matters, it has to be added back.

The most effective way to improve the result starts before editing, at the shoot: a plain background that contrasts clearly with the product, diffused light that softens shadows, and focus that keeps the product's edges sharp make the model's job much easier. In batch work, instead of opening every image, zoom in on the edges of a random sample to catch problem product types early, and check transparent or reflective products by hand.

Object removal: yes on the background, no on the product

Object removal is a strong tool for cleaning up a cable, dust or a stain on the backdrop, a hand that slipped into the frame or clutter in the background. Used on the product itself, the same tool shows the product as something it is not: erasing a scratch, a stitching flaw or a color difference means showing an image of a product the customer will not receive. In Türkiye, where TecnoNest is based, Article 61 of the Consumer Protection Law No. 6502 counts marketing communication through visual means as commercial advertising and requires it to be truthful and honest, and Article 7 of the Regulation on Commercial Advertising and Unfair Commercial Practices bans statements or images that could mislead consumers about a product's structure, composition, accessories or technical features.

The same line applies to images that belong to someone else. Removing a photographer's or supplier's watermark or rights information adds a separate problem to using the image without permission. Türkiye's Law No. 5846 on Intellectual and Artistic Works (Additional Article 4) prohibits removing or altering information about a work, its author or its rights holder without authorization, and Article 71 of the same law provides for imprisonment for this act; using AI to erase a watermark that carries such information falls under this ban. The safe route is to get permission to use the image, or to photograph the product yourself.

Upscaling and sharpening: the added detail does not come from the photo

AI upscaling is often used to bring a small image up to the size a sales channel requires. While increasing the pixel count, the model adds detail the image does not have; the result looks sharp to the eye, but nothing shows that the added texture is the product's real texture. Small print on a label, ingredient lists, barcodes, serial numbers and logos are the details that suffer most; the model can replace letters with shapes that merely look like letters.

That makes upscaling a fallback for when there is no other source. Asking the manufacturer or photographer for the original high-resolution file, or reshooting the product, always gives a more accurate result. On an upscaled image, always compare text and labels with the original.

Colorization and photo restoration: fine for archives, not for products

Colorization and restoration are useful for reviving old family and company photos and archive images. Neither belongs on a product image: colorization cannot know the product's real color, and restoration changes the product's real condition. With archive photos, too, the right practice is not to present a colorized or restored image as the original, and to say in its caption that it was edited.

Because face restoration redraws a blurred face, it should never be applied to images used for identity verification, such as passport-style photos, passport photos and ID card photos. For the same reason, an AI-generated professional headshot does not replace an ID photo.

What sales channels expect from a product image

Sales channels want product images to show the product as it is, and they put this in writing. Google Merchant Center publishes some of the most detailed rules; the images of products shown on Google Shopping are checked against them. According to Google's image link page:

• The image must show the exact item being sold. Placeholders, store logos, one image reused across different products, and drawings that do not show the physical product are not accepted. Each variant gets its own image, showing that variant's real color, pattern and material.
• Nothing extra on the main image. Prices, discounts, shipping information, calls to action such as "Buy", watermarks, copyright symbols, logos added after the shoot, barcodes, QR codes and borders are not allowed. The product should fill roughly 75 to 90 percent of the image.
• Size and format: From January 31, 2027, Google requires at least 500×500 pixels for all products and recommends about 1500×1500 pixels or larger. Images may not exceed 64 megapixels or 16 MB; JPEG, WebP, PNG, GIF, BMP and TIFF are accepted, and the file extension must match the real format.
• No upscaled images. Google's "Image too small" page explicitly says not to submit thumbnails or scaled-up images. On this channel, AI upscaling is not a fix for a size problem.
• AI-generated images keep their label. Google requires product images created with generative AI to keep the metadata tag that says so: the trainedAlgorithmicMedia value in the IPTC DigitalSourceType field.

The last rule is easy to miss in a workflow. Many tools that resize or compress images strip metadata to make the file smaller, so an AI-generated image can come out of that step without its label. Check that the tag is still present in the final file before it is published. The IPTC digital source type vocabulary uses separate values for an image created from scratch, an image edited with generative AI (object removal, for example) and a conventional correction such as sharpening; C2PA (Content Credentials), the open standard that records content provenance and edit history in the file, uses the same values.

Marketplaces publish their own image sizes and background rules in their seller panels and change them from time to time; figures circulating on third-party sites often contradict each other. The seller panel of the marketplace you sell on is the only reliable source, so check the rule there before preparing images.

There is also a new rule on AI in advertising: in Türkiye, paragraph 8 of Article 18 of the Regulation on Commercial Advertising and Unfair Commercial Practices, in force since August 1, 2026, requires that when AI or other software is used in an ad in a way that significantly affects consumers' economic behavior, this is disclosed in a clear, understandable and distinguishable way.

When AI photo editing is the wrong tool

• Edits that change the product itself: erasing a defect, changing the color, adding a part the product does not have.
• Photos that serve as evidence: damage, appraisal, delivery, return and warranty photos; an edited evidence photo has no evidential value.
• ID photos: passport and ID card photos; face restoration and AI headshots do not replace them.
• Images that belong to someone else: editing images you have no permission to use, or removing their watermark.
• Images with exact detail: technical drawings, labels, ingredient lists, barcodes and certificates; upscaling and sharpening can distort these details.

Photos of people and data protection

If a recognizable person appears in a photo (a model, an employee, a customer), the image is personal data under data protection laws such as the GDPR and Türkiye's KVKK, and uploading it to an editing service is processing that needs a legal basis; if the service processes data abroad, the transfer rules apply too. Türkiye's Personal Data Protection Board has fined a company that kept a former employee's photos on its social media account without explicit consent, and ordered the photos deleted.

A photo is not automatically biometric data, though. Türkiye's data protection authority, following GDPR Recital 51, states in its guide on special categories of personal data that a photo becomes biometric data only when it is processed with a specific technical method that allows a person to be uniquely identified or authenticated. Removing a background or a stain does not fall into that category; processing aimed at identifying a person by their face or matching faces falls under the rules for special categories of personal data. When choosing a service, check where images are processed, how long they are kept and whether they are used for model training; Google notes for its own product image tool, Product Studio, that quality reviewers may see inputs and outputs.

Batch processing: web panel, API and AI tools

For a handful of images, a web-based editing tool is enough. As a product catalog grows, uploading images one by one gives way to connecting the process to your systems: when a new product is added, its image has its background removed, is resized to the set dimensions and is saved as WebP automatically. That is what a REST API is for. Teams working with AI assistants such as Claude have a third option: connecting the image service through an MCP server as a tool the assistant can use, so that a request like "remove the background from the product images in this folder and crop them square" is carried out with the service's tools.

TheGlitch, the image processing service run by TecnoNest, offers all three routes in one service. Background removal, upscaling, colorization, photo and face restoration, and object removal run on AI models; basic operations such as cropping, resizing and format conversion use classical image processing. Images can be processed one at a time or in batches from the web panel, from your own systems through the REST API, and from inside AI tools through the MCP server. Current plans and prices are published on TheGlitch's own site; you can see TecnoNest's other AI products on our products page.

Frequently Asked Questions

How do you remove a photo's background with AI?

You upload the image to a background removal tool; the model produces a mask that separates the product from the background and either leaves the background transparent or fills it with a color you choose. Save the result as PNG or WebP to keep the transparency; JPEG does not support it. Zoom in on the edges of hair, fur, glass and reflective surfaces; a shoot with clear contrast between product and background is the step that improves the result most.

Does removing objects from a product photo make it misleading?

Not if the removed object is around the product, such as a cable on the backdrop, dust or a hand in the frame. If what is removed is a defect of the product, such as a scratch, a stain or a stitching flaw, the image shows the product as something it is not. In Türkiye, the Regulation on Commercial Advertising and Unfair Commercial Practices bans images that mislead consumers about a product's structure or technical features, and Google Merchant Center requires the image to show the exact item being sold.

Does AI upscaling really improve photo quality?

The pixel count goes up and the image looks sharper, but the added detail does not come from the photo; the model predicts it from patterns learned on similar images. Exact details such as label text, barcodes and logos can distort. Google Merchant Center explicitly asks merchants not to submit scaled-up images; where possible, use the original high-resolution file or reshoot the product.

Can AI-⁠edited product photos be used on Google Shopping?

Yes, if the image shows the product as it is and follows Merchant Center's size, format and "nothing extra on the main image" rules. If the image was created with generative AI, Google requires the file to keep the IPTC DigitalSourceType metadata tag that says so (trainedAlgorithmicMedia); check that the tools you use to resize or compress images do not strip it.

Does AI colorization give an old photo its real colors?

Not necessarily. A colorization model predicts a likely color for each region of a black-and-white photo; it is convincing with typical colors such as sky, leaves and skin, but cannot know the real color of a piece of clothing, a car or a sign. Do not present a colorized archive photo as the original; say that it was edited.

Is it legal to upload photos of people to an AI editing service?

A photo of a recognizable person is personal data, so uploading it is data processing that needs a legal basis under the applicable data protection law, and transfer rules apply if the service processes data in another country. According to Türkiye's data protection authority, which follows GDPR Recital 51, processing a photo is not automatically processing biometric data; a photo becomes biometric data only when it is processed with a specific technical method that allows a person to be uniquely identified or authenticated.

How do you remove backgrounds in bulk?

For a small number of images, a batch upload in a web panel is enough. If your product catalog grows regularly, connecting the process to your own system through a REST API means background removal, resizing and format conversion happen automatically whenever a new product is added. In batch work, check a random sample by hand, and review transparent and reflective products separately.