Low resolution product images can be repaired when the original captured the detail and something else destroyed it: a resize pipeline, a compression pass, a bad export. They cannot be repaired when the camera never recorded the detail. Upscaling adds pixels, not information. Sort the catalog on that one distinction, restore the files that pass it, and reshoot the ones that do not.
Many brands with more than two years of history have this folder. Product shots that came off a supplier listing, survived a platform migration, and now sit at 800 pixels wide with a filename nobody recognizes. They were fine when the site rendered them at 400. They are not fine now.
A study by eBay puts a number on what that costs. Across 6.8 million listings, listings with better photo quality were 4.5% more likely to sell. Here, “better photo quality” meant photos of 500 or more pixels on the longest side, with no added text or graphics (eBay photo tips, accessed September 2026). That is a vendor’s first-party figure rather than independent research, and eBay has an obvious interest in sellers uploading better photos. The sample size is what makes it worth quoting.
This covers what marketplaces require, why upscaling does less than its marketing suggests, and a triage rule for deciding which legacy files to restore, which to rebuild, and which to shoot again.
Key Takeaways
- Upscaling recovers pixels, not detail. A model can invent plausible texture where none was recorded. That is generation, not restoration, and on a product photo it is a returns risk.
- Amazon needs 500 to 10,000 pixels on the longest side, and 1,000 pixels or more turns on zoom. The main image needs a pure white background at RGB 255,255,255, with the product at 85% of the image (Amazon product image guide, accessed September 2026).
- Google Merchant Center is mid-transition. Google announced the change on April 14, 2026, and shows warnings for small images before a universal 500 x 500 minimum is enforced from January 31, 2027 (Google Merchant Center Help, accessed September 2026).
- Google requires AI-generated product images to keep their AI metadata. The IPTC DigitalSourceType tag must survive your export pipeline (Google Merchant Center Help, accessed September 2026).
- The editing model caps your output. FLUX.1 Kontext outputs around 1MP total, while Nano Banana Pro generates up to 4K (docs.bfl.ai and ai.google.dev, September 2026). Pick for resolution when resolution is the job.
- Repair in passes, upscale last. Clean, relight, then cut out, then enlarge. Upscaling first bakes the flaws in at a larger size. The DesignerBox image editor makes one edit after another, and the picture keeps its detail and resolution.
- Price the run before you start it. In DesignerBox, the model you edit with sets the cost, and you see it before the run.
What are low resolution product images?
Low resolution product images are files that fall below the minimum of the surface they appear on. A file also counts when enlarging it to fill that surface shows softness a shopper can see. The pixel count alone does not settle it. A sharp 900 pixel file often outperforms a soft 2,000 pixel one that was upscaled from 600.
Four different failures get filed under “low res”, and they have different fixes.
Undersized but sharp. The file is small and clean. Nothing was lost, but there was never much of it. This is the one case where enlargement works, because the detail is present and only needs more pixels to sit in.
Compressed. JPEG artifacts, blocking around high-contrast edges, color banding in gradients. The detail was captured and then partly discarded. Some of it is recoverable.
Soft at capture. Missed focus, camera shake, a cheap lens, or a phone shot in poor light. The detail was never recorded. No amount of processing brings it back, because there is nothing to bring back. Sharpening adds edge contrast and no detail. Our guide to when sharpening works and when to reshoot draws that line.
Damaged. Scratches on a scan, dust, water marks, a burned-in watermark or price flash from a marketplace listing. This is a repair job, not a resolution job, and it is the one that responds best to AI.
Sort the folder into those four buckets before you touch a tool. Two of them are worth your time and two of them are a reshoot.
What marketplaces require
These are the published minimums, taken from each platform’s own documentation. They are low, and the recommended sizes are what matter.
| Platform | Minimum | Recommended | Maximum | Frame fill |
|---|---|---|---|---|
| Amazon | 500 px longest side | 1,000 px or more turns on zoom | 10,000 px longest side | 85% |
| Shopify | Not published as a floor | 2048 x 2048 for square | 5000 x 5000, or 25 MP, under 20 MB | Not specified |
| Google Merchant Center | 500 x 500 from Jan 31, 2027 | 1500 x 1500 or above | 64 MP, under 16 MB | 75% to 90% |
| eBay | 500 px longest side | About 1600 x 1600 | 12 MB per upload | Not specified |
Sources: sellercentral.amazon.com, help.shopify.com, support.google.com/merchants and ebay.com. All four rows checked in September 2026.
Shopify is the outlier here, because it publishes no floor at all and caps media per product instead. If that is your platform, the Shopify limits that size the whole image job matter more than the resolution minimum. The delivered size Shopify serves matters more than either. PrestaShop builds every size from one master upload, so a soft master turns soft at every size, as PrestaShop product images explains.
Two things in that table are worth pulling out.
Amazon’s product image guide states the 500 pixel minimum, the 10,000 pixel maximum and the zoom threshold: images of 1,000 pixels or more on the longest side turn on zoom. The 1,600 pixel figure you may see elsewhere comes from Amazon’s 2021 clothing guide, which Amazon marks as informational.
Google Merchant Center is the one with a live deadline. Its attribute specification lists 500 x 500 as the minimum. Until January 31, 2027, Google still accepts images of at least 100 x 100 pixels for general products and 250 x 250 pixels for clothing (Google Merchant Center Help, October 2026).
Google announced the change on April 14, 2026 (Google Merchant Center Help, accessed September 2026). Merchant Center shows warnings for images below 500 x 500 before enforcement starts on January 31, 2027 (Google Merchant Center Help, accessed September 2026). If your feed carries legacy supplier images, that is a dated reason to audit them now rather than next year.
Why upscaling recovers pixels, not detail
An upscaler takes a small image and produces a larger one. It does that by predicting what should sit between the pixels you have, based on everything it learned from other images. When the source is sharp, those predictions are close to correct and the result is a clean enlargement. When the source is soft, the predictions are invention. Our guide explains how image upscaling works.
That distinction has consequences on a product photo that it does not have on a holiday snapshot. If a model invents a plausible weave on a fabric that was too soft to resolve, it has produced a texture your product does not have. A shopper who orders on that image and receives something else files a return, and the image did that. That risk is why the choice of upscaler matters, and we sort the tools by it in our Magnific AI alternatives comparison.
The rule is plain. Enlargement is safe when the detail exists and needs room. It is a fabrication risk when the detail is absent. The same tool does both, and it does not tell you which one it just did.
Our guide to why a supplier’s listing photo constrains everything downstream works through the same problem for video, where the failure is more visible because the image fills a phone screen. Our guide to AI video upscaling covers what an upscaler can and cannot fix in a clip.
Restore, rebuild, or reshoot
Here is the triage. Run each file against it before opening a tool. The tool will process all three categories, and you only see the difference after you have spent the credits. For a file that lands on the reshoot row, read what the replacement frame has to carry before you take the new photo.
| Condition of the original | Verdict | What you do |
|---|---|---|
| Sharp, small, clean | Restore | Enlarge. The detail is there and holds up. |
| Compressed, artifacts, but focus is good | Restore | Clean the artifacts first, then enlarge. |
| Scratched, dusty, watermarked, badly lit | Restore | Repair and relight. Resolution is not the problem here. |
| Soft at capture, missed focus, motion blur | Rebuild | Do not enlarge. Use a sharper photo of the same product as the source and derive the shot you need. |
| No usable photo of the product exists | Reshoot | Take one clean photo. Everything else derives from it. |
The rebuild row needs the most care. A soft photo is not a resolution problem and no upscaler fixes it, but you usually do not need to book a studio either. If any sharp photo of that product exists anywhere, even an awkward angle on a phone, that file becomes the source. You derive the angles, backgrounds and lighting from it. Our complete shot list for product angles covers what to derive once you have one good frame. How to fix product photo lighting covers which lighting faults survive a correction pass and which need the reshoot after all.
Reshoot only when nothing usable exists. A shoot costs real money, and what brands pay for a product shoot sets expectations before you book.
The repair passes, in order
Order matters, and the intuitive order is wrong. It is tempting to enlarge first, because it feels like the foundational fix. Enlarging first means every scratch, artifact and lighting flaw gets enlarged with it, and the later passes then have to work around damage that is now bigger. For a photo that needs no repair, how to edit product photos lists all eight editing steps in order.
- Repair the damage. Scratches, dust, blemishes, watermarks and burned-in text. In DesignerBox, an object removal step fills the marked area to match the surrounding pixels.
- Fix the lighting. Flat or mixed lighting reads as amateur at any resolution. A relighting step relights a product shot without a reshoot.
- Cut out the background. A clean edge is what most marketplace main images require, and it is easier to get on a repaired file than a damaged one. The White background photos template makes the cutout, and the Remove a background app does it for one photo, and what happens after the cutout covers the platform specs the replacement background has to clear.
- Enlarge last. Now the file is clean, the enlargement has good information to work from.
One change per pass. Starting the whole image again instead of correcting one thing at a time throws away the product accuracy you already established. The same principle sits behind making AI images that do not look generic AI.
Why the editing model caps your output
When you edit an image with an AI model, the output resolution is set by that model, not by your input file. Feed a 3,000 pixel photo into an editor that caps at 1MP and you get back something smaller than you started with.
The published caps vary widely.
| Model | Stated maximum output | Source |
|---|---|---|
| Nano Banana Pro | 1K, 2K or 4K | ai.google.dev, September 2026 |
| GPT Image 2 | 3840 x 2160 maximum, and OpenAI calls resolutions above 2560 x 1440 experimental | developers.openai.com, September 2026 |
| FLUX.2 family | Up to 4MP | bfl.ai, September 2026 |
| FLUX.1 Kontext | Around 1MP total | docs.bfl.ai, September 2026 |
Black Forest Labs lists Kontext under its previous-generation models and calls Kontext [pro] a legacy editing model, while FLUX.2 offers up to 4MP output (docs.bfl.ai, accessed September 2026). OpenAI also released GPT Image 2.5 on 8 September 2026, after GPT Image 2 (developers.openai.com, accessed September 2026). BFL says Kontext “excels at character consistency, even after multiple edits”, which is a different job from resolution.
The practical rule: when the edit is there to rescue a small file, pick the model by its output ceiling. In DesignerBox the model is one step in the workflow, and the model list shows the image models it runs. The image editor makes one edit after another, and the picture keeps its detail and resolution. So you can repair, relight, cut out and enlarge the same file, and the limit stays the ceiling of the model you chose.
What Google requires for AI-generated product images
If you restore a product image with AI and feed it to Google Shopping, there is a compliance step that has nothing to do with pixels. The same feed supplies AI shopping surfaces, and product images for AI shopping agents covers the image fields those feeds require.
Google requires that all images created using generative AI contain metadata indicating the image was AI-generated. Its example is the IPTC DigitalSourceType tag TrainedAlgorithmicMedia. Google instructs merchants not to remove embedded metadata tags (Google Merchant Center Help, accessed September 2026). Google also names compositeSynthetic and algorithmicMedia. IPTC’s own list adds compositeWithTrainedAlgorithmicMedia for images edited with generative AI, and algorithmicMedia means formula-made media, not generative AI (IPTC Digital Source Type, accessed September 2026).
That matters operationally because ordinary image pipelines strip metadata. A resize script, a CMS upload, a compression pass, or an export preset set to “minimal” will quietly remove the tag that Google is asking you to preserve. Check what survives your pipeline end to end, not what leaves the generator.
Whether you have to disclose AI imagery to shoppers is a separate question with a different answer per channel. What to check before shipping an AI product photo covers the accuracy side of the same decision. This is general information, not legal advice.
Cost of a restoration run
Price the run, not the file. A repair, a relight and an upscale are separate steps, and the model you run each one on sets its cost. In DesignerBox, you see the cost of each run before you start it, so you can price a folder first.
Each plan carries a monthly credit allocation. A repair pass over a folder of legacy files draws on it three times: once for the repair, once for the relight, once for the enlargement. Count the files before you pick the plan.
Uploading your own photos starts on the Pro plan. Upscaling and the image editor start on the Premium plan. Plans and credits are on the pricing page.
Dedicated upscalers meter use differently, often per image rather than from a shared credit pool, and a comparison of the tools built for ecommerce product imagery covers that trade-off. A single-purpose upscaler can be the better fit when enlargement is all you need.
Do the triage first. Then build the passes once, as a workflow. A saved workflow runs the same way on the next file, so the next supplier drop gets the same treatment.
Anyone can make an AI picture. Making hundreds that still look like your brand is the hard part. The brand rules sit in the workflow, so a repaired archive returns with one look. The repair, the relight, the cutout and the enlargement all happen in one account.
An upscaling app is the direct fix for a file that is small but sharp, and the comparison pages set DesignerBox next to dedicated tools. AI photo retouching tools compares the tools that handle the repair step itself. The AI product photography page shows the packshots, extra angles and styled scenes one photo becomes.
To test the passes on one file, start from a template.
FAQ
Can AI restore a blurry product photo?
Yes, if the photo is small but sharp, or compressed. AI restoration works well there because the detail was captured and only needs recovering. No, if the photo is soft because focus was missed or the camera moved. The detail was never recorded, and what a model produces in its place is invented. On a product image that invention is a returns risk, not a fix.
What resolution do product images need to be?
Amazon needs 500 to 10,000 pixels on the longest side, and 1,000 pixels or more turns on zoom. eBay requires at least 500 pixels and recommends about 1600 x 1600. Google Merchant Center recommends 1500 x 1500 or above, and enforces a universal 500 x 500 minimum from January 31, 2027 (sellercentral.amazon.com, ebay.com and support.google.com/merchants, September 2026). Shopify does not publish a floor, and says 2048 x 2048 usually displays best for square product images (help.shopify.com, September 2026).
Does upscaling a product photo reduce quality?
Upscaling does not reduce the quality of what was there, but it can add detail that was not. An upscaler predicts what belongs between existing pixels. On a sharp source those predictions are close to correct. On a soft source they are fabrication, and the enlarged file can misrepresent texture, stitching or finish that your product does not have.
Should I upscale before or after editing a product photo?
After. Repair damage, fix the lighting, then cut out the background, then enlarge. Upscaling first enlarges every flaw along with the image and forces later passes to work around damage at a bigger size. Enlarging last means the process has a clean file to work from.
Do AI-restored product images need to be labeled?
For Google Shopping, yes, at the metadata level. Google requires images created with generative AI to carry metadata indicating that, citing the IPTC DigitalSourceType tag TrainedAlgorithmicMedia, and asks merchants not to strip embedded metadata (support.google.com/merchants, September 2026). Watch your export pipeline, because resize and compression steps routinely remove metadata without warning.
Why did my edited image come back smaller than the original?
Because the editing model set the output size, not your input file. Published ceilings differ sharply: FLUX.1 Kontext outputs around 1MP while Nano Banana Pro generates up to 4K (docs.bfl.ai and ai.google.dev, September 2026). When the job is rescuing a low resolution file, choose the model by its output ceiling before anything else.
Is it cheaper to restore old product photos or reshoot them?
Restoring is usually cheaper when a usable original exists. In DesignerBox, the model sets the cost of a repair, relight and upscale. You see the cost before the run, so you can check a folder against your plan’s monthly credits. A reshoot is worth the money only when no sharp photo of the product exists at all, because a single clean frame is enough to derive the angles, backgrounds and scenes from.
Sources
- Amazon product image guide, including the 500 to 10,000 pixel range, the 1,000 pixel zoom threshold, pure white main-image background and 85% of the image (sellercentral.amazon.com, accessed September 2026)
- Shopify product media limits and the 2048 x 2048 square recommendation (help.shopify.com, accessed September 2026)
- Google Merchant Center image requirements and the generative-AI metadata requirement (support.google.com, accessed September 2026), and the image size change with enforcement from January 31, 2027 (support.google.com, accessed September 2026)
- IPTC Digital Source Type vocabulary (cv.iptc.org, accessed September 2026)
- eBay adding pictures guidance, about 1600 x 1600 recommended and 12 MB uploads (ebay.com, accessed September 2026), and the 4.5% likelihood-to-sell study across 6.8 million listings (export.ebay.com, accessed September 2026)
- FLUX.2 and FLUX.1 Kontext output resolution (bfl.ai and docs.bfl.ai, accessed September 2026)
- Nano Banana Pro resolution tiers (ai.google.dev, accessed September 2026)
- GPT Image 2 size constraints (developers.openai.com, accessed September 2026) and the GPT Image 2.5 release (developers.openai.com, accessed September 2026)
- DesignerBox plans, credits and license gating (DesignerBox pricing, designerbox.ai/pricing, September 2026)
A note on the 4.5% likelihood-to-sell figure: eBay publishes it on its own seller guidance as its own study, with its method stated, so it is quoted here as eBay’s figure.
Marketplace image requirements verified from each platform’s own documentation, and model output resolutions from each provider’s own developer documentation, as of September 2026. Individual results vary.
