Where AI holds up for product images, and where it fails marketplace rules

Four jobs get handed to AI in a product image pipeline: cutting the background out, writing alt text, tagging the image with attributes, and enlarging a photo that is too small. They get treated as one capability. They are four, each with its own failure mode, and the marketplaces they feed check images against rules specific enough to reject a file for a band of grey pixels along one edge.

We have run these tools over catalogs from supplier photos through to live listings on Amazon.in, Flipkart and Meesho. The pattern repeats. The AI step is fast, the output looks right at thumbnail size, and the rejection arrives from a rule the tool never knew about.

Background removal is dependable on hard edges and unreliable on everything else

A solid object with a clean outline, such as a box, a mug or a shoe, comes out of a removal model cleanly almost every time. The failures cluster around edges the model has to guess: hair and fur, mesh and lace, glass and transparent bottles, steam, drop shadows, and any white or pale product shot against a light background. The model either eats part of the product or leaves a soft halo of the old background around it.

The marketplace rule that matters here is the main image background. Amazon's requirement is pure white, RGB 255,255,255, and Flipkart and Meesho ask for white or plain backgrounds across most categories. A halo, a leftover shadow, or a transparent PNG that renders grey in some viewers all fail it. Flatten every cutout onto a solid white layer, export the main image as JPEG, and inspect the edges at 200 percent zoom before you trust the file. The Marketplace Image Checker samples the border pixels and reports how many are near-white; a low figure means the edge needs a second look, and it flags transparency outright for platforms that forbid it.

A cutout leaves the product small in the frame

Removing a background does not change the canvas. The product keeps whatever padding the photographer gave it, and on a plain white field that padding becomes empty space. Amazon asks for the product to fill 85 percent or more of the main image, and a cutout straight from the removal tool rarely does. Crop to the product's bounding box, add an even margin, and resize so the longest side stays at or above 1,000 pixels, which is the threshold for zoom. Under 500 pixels on the longest side the image is not accepted at all.

Crop after removal, then resize after the crop. Reversing the order either shrinks the product before you have found its edges, or crops a low-resolution file and lands under the minimum.

Upscaling passes the dimension check and misrepresents the product

An upscaler does not recover detail. It invents detail that would be plausible at the new size. On a holiday photograph that is harmless. On a product photo it draws stitching that is not on the garment, sharpens a texture into one the fabric does not have, and turns label text into letter-shaped marks that read as nothing. The image now clears the minimum dimension and shows a product that does not exist.

Use upscaling for modest enlargements of images with no fine text, and never on a photo where a label, an ingredient panel or a size chart is legible. For those, request the original from the supplier or reshoot. A rejected image costs an hour. A delivered product that does not match its picture costs the return.

Auto-tagging knows what a thing is and guesses what it is for

Vision models tag reliably at the level of object and colour: shoe, sneaker, red, lace-up. They are poor at material, gender, size, brand and intended use, and they will supply a value anyway. A tag of 'leather' on a synthetic shoe or 'women's' on a unisex one becomes a filter facet that lies to buyers, and then an attribute that ends up in generated copy.

Treat tags as suggestions for a person to confirm, never as facts that flow straight into a filter or a feed. The safe use is triage: grouping ten thousand supplier images by what they show so a human works through them in a sensible order, and catching the image uploaded against the wrong SKU because the model tagged a kettle on a toaster listing.

AI alt text describes the picture, and the buyer needs the product

Ask a vision model for alt text and you get 'a red sneaker on a white background'. That is an accurate caption and a useless alt text. A screen reader user on the product page already knows they are looking at the red sneaker listing, because the title told them. Google Images has the same problem from the other side: the alt text carries none of the terms a buyer would search.

Useful alt text for a product image identifies the product, the variant, and what this particular view shows that the other images do not. Brand and product name, the variant attribute where it is visible, the angle or context, and the detail the image was taken to show. Keep it under about 125 characters, since some screen readers stop announcing around there, and leave out 'image of' and 'photo of', which the reader already announces.

Weak:   A red sneaker on a white background.
Better: Nova Runner 2 in crimson, side view showing the reflective heel strip.

Weak:   Kitchen appliance on a counter.
Better: Vayu 1.5 L electric kettle in steel, lid open to show the concealed element.

You can still use the model for this. Give it the product title, the variant attributes and the list of images already on the listing, and ask what this image shows that the title does not. That question produces a usable draft. The bare image never does.

Added text and watermarks are a pipeline problem, not a model problem

Amazon.in, Flipkart and Meesho all reject main images that carry text, logos, badges, watermarks, borders or collages. AI rarely adds these on its own. Pipelines do. A caption bar made for social media gets reused as the listing image, an offer badge is composited in bulk, or a supplier watermark survives the removal step because it sat on the product itself. Generated lifestyle backgrounds on secondary images are a newer form of the same risk, since they sometimes render shelf labels and signage as letter-like shapes that a reviewer reads as text.

Can you use AI to prepare product images for Amazon or Flipkart?

Yes for background removal on hard-edged products and for triaging images by content, with a check of edges, white background, frame fill and dimensions after each step. Be cautious with upscaling, which invents detail, and with auto-generated alt text and tags, which describe the picture instead of identifying the product and need a person to confirm.