Tecof • September 15, 2026
AI Product Image Optimization: Background Removal
In Brief
Background removal is the job of separating the object in a product photo from its background and turning it into a transparent or flat-backed image. What used to take five to ten minutes per image in Photoshop is now a step a model completes in seconds. But the real task is not cleaning one image; it is getting an entire catalog onto a single standard with the same framing, the same ground, the same padding and the same file size. As of 2026 the question is no longer "can AI remove the background" but "at what quality threshold, in how many minutes and with how many people do we get a thousand images to production standard".
Tuesday morning, 09.15. The new season has 840 products and the supplier images do not match each other: 310 are white-background studio shots, 260 sit on grey fabric, 190 were taken on a desk with a phone, and 80 show another brand's label. Listed on product pages, the catalog looks scattered; of six products side by side on a category page, three are white, two grey, one on wood. Conversion is below the same category last season and nobody can say why.
The cause is usually not image quality but inconsistency. Comparing products on a category page, the eye reads a difference in background as a difference in product; the grey-background item feels "cheaper", the one on wood feels second-hand. Removing the background is only the first step; the real gain hides in the common framing rule applied afterwards.
Why Product Images Are the Cheapest Conversion Lever
What a user processes in the first second of an e-commerce page is not text but image. They decide what the product looks like before reading the price or the description. That is why improving images pays off faster than improving copy: writing a new description takes half an hour and its effect is gradual, standardising an image takes minutes and shows up on day one.
Consistency beats individual quality
A hundred images do not all need to be perfect; they need to be the same. Same ground colour, same product-to-frame ratio, same shadow behaviour. Lined up on a category page the eye sees a single rhythm and the comparison happens on the product itself. In a mixed catalog the user unconsciously equates image quality with brand quality.
Images are part of page speed too
Removing a background usually shrinks the file, but saving as PNG does the opposite: a catalog using PNG for transparency ends up with 1.5-3 MB files per image. Six images on a product page is a 10 MB load on its own. Use WebP if you need transparency and plain white-background JPEG if you do not. In our piece on raising conversion rates we explained why page speed comes first.
Every image is a data record
An image's filename, alt text and dimensions are signals for search engines too. "IMG_2831.jpg" says nothing; "black-leather-ankle-boot-side-view.webp" works in both image search and page context. Wire naming into the same pipeline and every new product inherits the standard automatically.
How Background Removal Works
You do not need the technical detail, but the basic logic helps explain why output sometimes breaks.
Segmentation and the mask
The model decides, for every pixel, whether it belongs to the product or the background, and produces a "mask". The mask is a black-and-white stencil: white stays, black is removed. Output quality depends on how precise the mask's edges are.
- Mask: the black-and-white stencil deciding which pixels stay. A hard-edged mask produces jagged results on hair, fur and mesh.
- Alpha matting: the refinement that gives edge pixels partial transparency. This is the step that makes the difference on hairy, furry, transparent or mesh-textured products.
- Halo: the thin light line left around the product after removal. It shows up most on dark-background shots moved onto white.
- Spill: background colour bleeding into the product's edge. On green or blue screen shots the product edge picks up a green cast.
Which products are easy and which are hard?
Sorting your catalog by difficulty before building the pipeline sets the right expectations. Products with a simple silhouette, opaque and high in contrast come out publish-ready almost every time. Transparent, shiny, furry products or ones the same colour as the background require human review.
| Product type | Difficulty | Typical defect | Recommended flow |
|---|---|---|---|
| Shoes, bags, boxed goods | Low | Almost none | Fully automatic |
| Apparel (on a mannequin) | Medium | Hair and hanger edges | Automatic + sampled review |
| Jewellery, eyewear, glassware | High | Transparent areas erased | Automatic + item-by-item review |
| Fur, knitwear, mesh | High | Jagged edges, halo | Alpha matting + manual fix |
| White product on white | High | Product edge disappears | Reshoot or manual mask |
Removal is not enough: framing
After the background is gone the product still sits somewhere random inside the frame. The production standard is framing, not removal: crop the product, place it in a square or 3:4 frame, leave equal padding on every side, set the canvas on the required ground. Skip automating this step and the catalog stays scattered, only with white backgrounds.
Marketplaces and Your Own Site: Different Rules
A store selling in Turkey usually lists the same product in three places, and the three have different image expectations. Producing one "main image" and sending the same file everywhere breaks a rule in at least one of them.
Marketplace expectations
Marketplaces such as Trendyol and Hepsiburada generally require a plain white background on the main image, a set amount of padding, and no text, logo, watermark or border. These rules change from time to time; when building the pipeline take the current image guide in the relevant seller panel as your reference rather than assuming they are fixed. A violation means the product is not published or quietly goes inactive later, and you usually notice when sales drop.
Your own site gives you more room
On your own store the main image does not have to be white. A light grey close to your brand colour separates the product better than white and reads softer across the page. The only rule here is the one you set yourself, and if you do not write it down nobody will remember it in three months.
| Where it is used | Ground | Ratio | Format | Target size |
|---|---|---|---|---|
| Marketplace main image | Plain white | 1:1 | JPEG | 300-600 KB |
| Own site product page | Light grey or white | 1:1 or 3:4 | WebP | 80-150 KB |
| Category thumbnail | Same as main | 1:1 | WebP | 20-40 KB |
| Social media image | Brand colour | 4:5 / 9:16 | JPEG | 150-300 KB |
| Ad creative | Scene or flat | 1:1 / 4:5 | JPEG | 150-300 KB |
One source, many outputs
The right setup is this: produce one transparent "master" file, then derive every channel version from it. If the master changes, all derivatives are regenerated. Without a master you end up recropping the white-background JPEG you made for a marketplace into a social post and losing quality on the way.
How Do You Push a Thousand Images Through?
A method that works on one image collapses on a thousand, because the problem is not model accuracy but error rate multiplied by scale: 95 percent success means fifty broken images out of a thousand, and finding those fifty one by one can take longer than doing the lot by hand.
Sort first, process second
Before bulk processing starts, split the catalog into three buckets: will pass automatically, needs sampled review, needs item-by-item review. The split can be made by product category and takes five minutes. Start without it and the hard items lost among the easy ones ship broken.
Break the pipeline into steps
- 1. Input check: do not process anything under 1,000 pixels, list it instead. You cannot make a large image out of a small one.
- 2. Removal: mask generation and transparent master save.
- 3. Crop and frame: centre the product, apply padding as a percentage.
- 4. Ground and derivatives: apply ground, ratio and format per channel.
- 5. Naming and alt text: save with the product code and a descriptive name.
- 6. Inspection: automated checks plus human sampling.
The reason to keep steps separate is this: when one breaks, you rerun only that one. Merged into a single command, a naming mistake forces you to reprocess a thousand images from scratch. If product coding is messy the pipeline jams at the very start; a disciplined SKU structure is a prerequisite for the image line too.
Think of images alongside product data
Image production is not a standalone job but part of product entry. Generate the description, title and images in the same batch and the product is publish-ready in one pass. The acceptance-criteria logic we described in the piece on bulk product descriptions applies here too: you check output against a measurable list, not against "I liked it". Running catalog, orders and media from one place requires connections between systems to be in place; otherwise the manual upload after generation eats the whole gain.
Quality Control: What Are You Measuring?
The answer to "is this image good" varies by person. A five-item checklist is enough to make it measurable, and most of that list can be checked automatically.
What can be checked automatically
- Ground purity: do the corner pixels deviate from the expected ground colour?
- Product fill: what percentage of the canvas does the product occupy? Outside the target range means framing is off.
- File size: does it exceed the channel's ceiling?
- Resolution and ratio: are they at the expected values?
- Empty area share: if the transparent pixel share is far above expectations, part of the product may have been erased.
What a human has to look at
Edge quality, halo, a clipped part of the product and shadow realism cannot be measured reliably by machine. Sampling is enough: open ten out of every hundred images at full size and look. If more than two of ten show defects, reject that batch and change the settings; one or zero and the batch passes.
| Check | Method | Threshold | If it fails |
|---|---|---|---|
| Ground purity | Automatic | 100% match | Queue the image |
| Product fill | Automatic | 75-90% | Rerun framing |
| File size | Automatic | Below channel limit | Increase compression |
| Edge quality | Sampling | At most 1 defect in 10 | Reject the batch |
| Missing part | Sampling | Zero tolerance | Manual mask |
Count the cost of going back
The cost of a broken image is not the time to fix it but the time it stays live. Changing the image on a published marketplace listing can take a day because of approval. Loosening thresholds and saying "we'll fix it later" costs more than tightening them and rejecting the batch.
Copyright, Trademarks and Data Protection
Image processing looks like a technical job but it has a legal side, and that side is usually considered last.
The right to use supplier images
Do not assume you have the right to use images that came from a supplier. Unless the dealership or supply agreement explicitly grants image usage, processing that image into your own catalog can cause problems. Removing a background does not make the image "a new work".
Another brand's logo and the model's face
Two things slip through bulk processing most often: another brand's label visible in frame, and a model's recognisable face. The second is personal data under KVKK; without a usage consent from the model, publishing the image becomes risky. Adding "is there a face" and "is there text" checks to the automated flow catches both before they go live.
AI-generated scene backgrounds
Removing a background and dropping in an AI-generated scene is technically easy, but it becomes misleading if it shows a context the product is not actually in. A room for furniture or a bathroom scene for cosmetics is generally acceptable; a scene that misrepresents the product's size, colour or contents is not. Consumer law expects images to represent the product accurately, and your return rate will confirm it.
Building an Image Line in Thirty Days
This work usually lives on one person's computer, in one folder, as one habit. When that person leaves, the standard goes too. The timeline below turns scattered work into a flow the company owns.
Days 1-7: write the standard
On a single page, decide: ground colour, product fill ratio, padding, aspect ratio, format, file size ceiling and naming rule. No images are processed this week. Bulk processing done before the standard is written gets redone in month two.
Days 8-14: try it on fifty images
Pick fifty products across easy, medium and hard categories and run the pipeline end to end. How many passed untouched, how many were fixed by hand, how many were rejected? Those three numbers let you estimate the effort for a thousand. A pipeline below 80 percent on fifty images does not go to production.
Days 15-21: wire up channel derivatives
Generate marketplace, site, category and social derivatives from the master automatically. This week's output is not images but a working flow that produces five outputs from one file. Wire the upload step in here too; every step left manual jams by month three.
Days 22-30: migrate the catalog and name an owner
Push the rest of the catalog through in batches, sampling every batch. By month end you should have written down: the standard document, the pipeline steps, the acceptance thresholds, the last measured pass rate, and the owner. An ownerless pipeline degrades within three months and nobody notices.
Here is the job for tomorrow morning: open your own category page, put the twelve product images visible in the first screen side by side, and count how many different treatments you have in ground colour, product size and padding. If the number is greater than one, you have just written the first line of your standard document. On an e-commerce setup where catalog, images and product data run in one flow most of these steps arrive ready-made and you only define the standard.
Frequently Asked Questions
Do I need a paid tool for background removal?
At small volumes free tools do the job. What paid tools add is not single-image quality but bulk processing, API access and consistency. Under a hundred images a month, free is fine; over a thousand, a tool with an API pays for itself in the first month.
Transparent PNG or white-background JPEG?
Keep your master transparent and your published file not. Transparent PNGs are large and behave unpredictably in dark mode. Publish white or grey-background WebP and keep the transparent master only for generating derivatives.
How long should one image take?
In an automated pipeline removal and derivative generation take a few seconds per image; a thousand finish in under an hour. The real time goes into review and correction. A realistic plan allows one day of bulk processing and two days of inspection per thousand images.
Can AI upscale low-resolution supplier images?
Upscalers invent detail; on a product image that means a seam that does not exist or a mangled logo. Rather than pushing an 800-pixel image to 2,000, ask the supplier for the original or reshoot. Upscaling is for closing a small gap, nothing more.
Should I remove the shadow or keep it?
Marketplace main images generally want a clean shadowless result. On your own site a soft contact shadow stops the product looking like it is floating and reads more real. Decide per category and write it into the standard; do not let one category hold both shadowed and shadowless images.
Can I use an AI-generated scene as the main image?
Marketplaces expect the main image to show the product plainly and accurately, and a generated scene is generally not suitable there. It has a place in the second and later images. On your own site you have more freedom, but the image must not misstate the product's size or colour.
How many people does this take?
Once the pipeline is built, half a day a week from one person covers a thousand-product catalog. Setting it up takes two: someone to define the standard and someone to build the flow. They can be the same person, but they are two different jobs and not done at once.
Does changing images hurt SEO?
If filenames and URLs change there can be a temporary dip in image search. That is why you set the naming rule once during a bulk migration and never change it again. Fix the alt text in the same pass and the short-term loss is more than recovered in the medium term.
Should I migrate the old catalog or start with new products?
Start with new products and migrate the old catalog by sales volume. Migrating the top hundred revenue products delivers most of the benefit of migrating everything and finishes in a week.
How should I archive the output?
Store master files by product code and do not archive derivatives; a derivative can always be regenerated. Archiving them multiplies storage cost and creates the question of which file is current.