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 in the first place. Upscaling adds pixels, not information. Triage the catalogue on that distinction first, then fix only the files that pass it.
Every brand with more than two years of history has 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 recognises. They were fine when the site rendered them at 400. They are not fine now.
eBay’s own research puts a number on what that costs. Across a study of 6.8 million listings, listings with better photo quality were 4.5% more likely to sell, where “better photo quality” meant photos measuring 500 or more pixels on the longest side, with no added text or graphics (export.ebay.com, July 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 actually 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 requires 500 to 10,000 pixels on the longest side, a pure white main-image background at RGB 255,255,255, and the product filling 85% of the frame (sellercentral.amazon.com, July 2026).
- Google Merchant Center is mid-transition and warnings started this month. A universal 500 x 500 minimum becomes mandatory on January 31, 2027 (support.google.com/merchants, July 2026).
- Google requires AI-generated images to keep their provenance metadata. The IPTC DigitalSourceType tag must survive your export pipeline (support.google.com/merchants, July 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, July 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 honest tier floor is Premium. Importing your own photos starts at Pro, and the upscaler and relight start at Premium at $75 a month.
What counts as a low resolution product image?
A product image is too low resolution when it falls below the minimum of the surface it has to appear on, or when enlarging it to fill that surface exposes 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, there was simply never much of it. This is the one case where enlargement genuinely works, because the detail is present and only needs more pixels to sit in.
Compressed. JPEG artefacts, blocking around high-contrast edges, colour 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.
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 actually require
These are the published minimums, taken from each platform’s own documentation. They are lower than most brands assume, and the recommended sizes are what actually matter.
| Platform | Minimum | Recommended | Maximum | Frame fill |
|---|---|---|---|---|
| Amazon | 500 px longest side | Not published as a figure | 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 | 800 to 1600 px longest side | 7 MB per upload | 80% to 90% |
Sources: sellercentral.amazon.com, help.shopify.com, support.google.com/merchants and ebay.com, all July 2026.
Two things in that table are worth pulling out.
Amazon’s guide states the 500 pixel minimum and the 10,000 pixel maximum, and it does not publish a zoom activation threshold. The widely repeated figure for that threshold appears only in seller forum posts, which contradict each other. It is left out here rather than stated as fact.
Google Merchant Center is the one with a live deadline. Its attribute specification lists 500 x 500 as the minimum, while the troubleshooting documentation confirms currently enforced floors of 100 x 100 for general products and 250 x 250 for apparel.
The universal 500 x 500 minimum becomes mandatory on January 31, 2027, and warnings for images below it began appearing in the Needs attention tab in July 2026 (support.google.com/merchants, July 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.
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.
This is why the rule is worth stating plainly rather than softening. 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.
Restore, rebuild, or reshoot
Here is the triage. Run each file against it before opening a tool, because the tool will happily process all three categories and only tell you the difference after you have spent the credits.
| Condition of the original | Verdict | What you do |
|---|---|---|
| Sharp, small, clean | Restore | Enlarge. The detail is there and holds up. |
| Compressed, artefacts, but focus is good | Restore | Clean the artefacts 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 middle row is the one brands get wrong. 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 and the angles, backgrounds, and lighting get derived from it. Our complete shot list for product angles covers what to derive once you have one good frame.
Reshoot is the honest answer only when nothing usable exists. It costs real money, and what brands actually pay for a product shoot sets expectations before you book.
The repair passes, in order
Order matters, and the intuitive order is wrong. Most people enlarge first because it feels like the foundational fix. Enlarging first means every scratch, artefact and lighting flaw gets enlarged with it, and the later passes then have to work around damage that is now bigger.
- Repair the damage. Scratches, dust, blemishes, watermarks and burned-in text. In DesignerBox this is Retouch, which brushes over the area and fills it to match the surrounding pixels. An edit costs 5 credits.
- Fix the lighting. Flat or mixed lighting reads as amateur at any resolution. Spotlight relights a product shot to studio standard 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. Background Remover produces the transparent file.
- Enlarge last. Now the file is clean, the enlargement has good information to work from. Basic upscale is 3 credits, Magnific upscale is 10.
One change per pass. Rerolling the whole image instead of correcting one thing at a time throws away the product accuracy you already established, which is the same principle behind making AI images that do not look generic AI.
Why the editing model caps your output
This is the part that surprises people. 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 by a factor of four.
| Model | Stated maximum output | Source |
|---|---|---|
| Nano Banana Pro | Up to 4K, 4096 x 4096 square | ai.google.dev, July 2026 |
| GPT Image 2 | 3840 px maximum edge, though its own docs call output above 2K experimental | platform.openai.com, July 2026 |
| FLUX.2 family | Up to 4MP | docs.bfl.ai, July 2026 |
| FLUX.1 Kontext | Around 1MP total | docs.bfl.ai, July 2026 |
Black Forest Labs now positions Kontext as previous generation and recommends FLUX.2 for new work, citing up to 4MP output against Kontext’s roughly 1MP (docs.bfl.ai, July 2026). Kontext remains strong at character consistency across multiple edits, which is a different job from resolution.
The practical rule: when the whole point of the edit is rescuing a small file, pick the model by its output ceiling. DesignerBox exposes the choice at the model catalogue rather than picking one house model for every job. Editing passes in DesignerBox hold detail and resolution, so the constraint is the ceiling you chose, not accumulated loss across passes.
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.
Google requires that all images created using generative AI contain metadata indicating the image was AI-generated, giving the IPTC DigitalSourceType tag TrainedAlgorithmicMedia as its example, and instructs merchants not to remove embedded metadata tags (support.google.com/merchants, July 2026). Related IPTC values include CompositeSynthetic and AlgorithmicMedia.
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, and what to check before shipping an AI product photo covers the accuracy side of the same decision.
What restoring a catalogue costs
In DesignerBox, credits map to operations. A repair or edit pass is 5 credits, a basic upscale is 3, and a Magnific upscale is 10. A damaged file that needs a repair, a relight and an upscale runs about 13 credits.
At that rate the plan allocations work out as follows.
| Plan | Price | Credits | Roughly, files repaired |
|---|---|---|---|
| Free | $0 | 112 | 8 |
| Basic | $15/mo | 500 | 38 |
| Pro | $35/mo | 1,000 | 76 |
| Premium | $75/mo | 2,500 | 192 |
| Ultra | $200/mo | 8,000 | 615 |
Two limits matter more than the credit maths, and they are easy to miss.
Importing your own photos starts at Pro. Restoration is by definition work on a file you already have, so Pro at $35 a month is the entry point for this job at all.
The upscaler and relight start at Premium. Hyper realism, edit, crop and zoom, relight and the upscaler are all Premium features. If your workflow needs enlargement and relighting rather than only inpainting, the real floor is Premium at $75 a month, not Basic.
Dedicated upscalers price differently, on image counts rather than mixed credit pools, and a comparison of the tools built for ecommerce product imagery covers that trade-off. A single-purpose upscaler can be the cheaper buy when enlargement is genuinely all you need.
For a full catalogue, Commerce Studio is where restored files feed the rest of the set, and the app list shows which pass each app handles. Current plan details are on the pricing page.
FAQ
Can AI actually restore a blurry product photo?
It depends on why the photo is blurry. If the image is small but sharp, or compressed, AI restoration works well because the detail was captured and only needs recovering. 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 requires 500 to 10,000 pixels on the longest side, eBay requires at least 500 pixels and recommends 800 to 1600, and Google Merchant Center recommends 1500 x 1500 or above with a universal 500 x 500 minimum becoming mandatory on January 31, 2027 (sellercentral.amazon.com, ebay.com and support.google.com/merchants, July 2026). Shopify does not publish a floor but recommends 2048 x 2048 for square product images (help.shopify.com, July 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 labelled?
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, July 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, July 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 cheaper whenever a usable original exists. In DesignerBox a repair, relight and upscale is around 13 credits per file, so Pro at $35 a month covers roughly 76 files. 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 requirements, including the 500 to 10,000 pixel range, pure white main-image background and 85% frame fill (sellercentral.amazon.com, July 2026)
- Shopify product media limits and the 2048 x 2048 square recommendation (help.shopify.com, July 2026)
- Google Merchant Center image_link specification, enforcement transition to a universal 500 x 500 minimum on January 31, 2027, framing guidance and the generative-AI metadata requirement (support.google.com/merchants, July 2026)
- eBay picture policy and photo guidance, including the 4.5% likelihood-to-sell study across 6.8 million listings (ebay.com and export.ebay.com, July 2026)
- FLUX.2 and FLUX.1 Kontext output resolution (docs.bfl.ai, July 2026)
- Nano Banana Pro resolution tiers (ai.google.dev, July 2026)
- GPT Image 2 size constraints (platform.openai.com, July 2026)
- DesignerBox credit costs, plan allocations and feature gating verified against live product configuration, July 2026
A note on one widely repeated statistic: the 4.5% likelihood-to-sell figure is frequently attributed to Cornell Tech research. That attribution does not hold up. The figure is eBay’s own internal study, published on eBay’s seller guidance with its methodology stated, and the citation chain behind the Cornell attribution leads back to eBay rather than to any university research. It is quoted here as what it is.
Marketplace image requirements verified from each platform’s own documentation, and model output resolutions from each provider’s own developer documentation, as of July 2026. Amazon’s zoom activation threshold is omitted because it does not appear in Amazon’s current published guide. Individual results vary.
