Ecommerce image editing is the work between a raw product photo and a listing that a marketplace will accept and a buyer will trust: pure white backgrounds, colour correction, shadows, detail and lifestyle images, colourways. This guide runs that work on one furniture SKU with Visualizee, an AI rendering tool that edits a product photo while keeping the product itself unchanged; Vizzy is its in-app assistant. We publish what vendor pages leave out: the marketplace specs with sources, the four rejection criteria we review against, every attempt including the failures, and the number that decides whether any of it is worth doing, the cost per accepted image.
Quick answer: A listing needs a main image on pure white (RGB 255) filling 85% of the frame, plus detail, lifestyle and variant images at 1,600-2,048 px. Outsourced editing costs $0.25-$0.95 per task per image with a 24-hour turnaround. AI costs about $0.27 per render, but in our measured run only 4 of 9 attempts passed review, so an accepted image cost $0.61 in renders and about $5.11 with review time included. Every job that kept the camera angle was accepted within two attempts; a new camera angle failed 3 times out of 3. The run used a rendered stand-in for a phone photo, so treat the acceptance rate as a best case.
What Does Ecommerce Image Editing Include?
Ecommerce image editing covers everything done to a product photo after capture so it meets a channel's spec and sells the product: background removal, colour and exposure correction, shadows, retouching, resizing, and the production of secondary images. Amazon recommends at least six images per listing; for a small furniture or home-goods brand the core set per SKU is five: a main image on white, a second angle, a detail, a lifestyle scene and one image per colourway.
Buyers lean heavily on those files. In Salsify's 2025 consumer research 77% of shoppers rated product images and video extremely or very important, and 42% named missing or low-quality images as a reason they abandoned a purchase. The same research found 71% had returned a product because the product information was inaccurate. The NRF and Happy Returns put 2025 US retail returns at 15.8% of sales, $849.9 billion, and 19.3% for online sales. An edited image that misrepresents the product invites exactly that kind of return, which is why this guide treats accuracy as the first editing requirement and aesthetics as the second.
What Are the Image Requirements for Amazon, Shopify, Google, Etsy and eBay?
Every major channel wants a clean, accurate, high-resolution image, and the strictest rule is Amazon's main image: pure white, product at 85% of the frame. Editing to the strictest spec first means one master file serves every channel. All figures below come from the platforms' own pages, checked on September 21, 2026.
| Channel | Size | Background and content rules |
|---|---|---|
| Amazon | 500-10,000 px on the longest side; more than 1,000 px for zoom; at least six images recommended | Main image on pure white (RGB 255, 255, 255), product fills 85% or more of the frame |
| Shopify | Up to 5000 x 5000 px and 20 MB; 2048 x 2048 px square "usually displays best" | Theme-dependent; square keeps grids consistent |
| Google Merchant Center | At least 500 x 500 px for all products from January 31, 2027; 1500 x 1500 or larger recommended; max 64 MP and 16 MB | No promotional overlays, watermarks or borders; must accurately display the entire product; generative-AI images must keep their IPTC metadata |
| Etsy | 2,000 px or more on each side; first photo landscape or square | sRGB; large files may fail to upload |
| eBay | 500 px minimum on the longest side, about 1,600 px recommended, up to 24 photos | No borders, added text, artwork or watermarks; photos must accurately represent the item |
Two rules in that table are new enough to catch teams out. Google's 500-pixel floor for every product takes effect on January 31, 2027, replacing the old lower minimums, so audit legacy images under 500 pixels before that date. And Google's AI-generated content policy requires that images created with generative AI keep their IPTC DigitalSourceType metadata; an export step that strips metadata to save a few kilobytes now breaks compliance. Amazon's equivalent rule is narrower: since June 2026 it asks sellers to tag photorealistic AI-generated people, as eWeek reported, so a furniture packshot with nobody in it needs no tag. Every channel keeps the older rule on top: the image has to represent the product accurately.
What Should You Reject? Four Criteria for AI-Edited Product Images
Reject any image where the product differs from the source photo in label, construction, finish or proportion, and check at 100% zoom, not at thumbnail size. The GS1 Product Image Specification, the standard retailers exchange product images under, sets the same bar for retouching: it must be undetectable at a minimum of 100% magnification, with no colour casts, no blown highlights, and shadows that never hide text or logos. GS1 also draws the line this guide depends on: an image that was "significantly altered", not merely retouched, counts as a rendered image. An AI edit that changes the product has stopped being a photo of it.
| Criterion | What to compare against the source | Typical AI failure |
|---|---|---|
| 1. Label and text | Spelling, typeface, position on the product, orientation | Misspelt or "almost" letters; label moved to a different seam or face |
| 2. Construction | Joint type and count, number of slats, legs, handles, stitching lines | Joints smoothed away; slats added, removed or turned into spindles; extra parts |
| 3. Finish | Species and grain direction, sheen, fabric texture, colour against a reference | Oak drifting to pine or walnut; boucle flattened to felt; colour warmed by the scene |
| 4. Proportion | Height to width, seat and cushion thickness, leg rake, camera angle | Product stretched to fill the frame; mirrored view; thicker, "comfier" cushions |
A fifth check belongs to the channel, not the product: measure the spec. Every packshot in the run below looked white on screen. Measured as the darkest pixel in a 40-pixel border, they came back between 189 and 246, and not one was a clean 255. The same goes for frame fill: both accepted packshots filled 78% of the frame by width, short of Amazon's 85%. Neither problem is visible at a glance, and both take a minute to fix once measured.
The Measured Run: One Chair, Nine Attempts, Four Accepted Images
One disclosure first: the source image in this run is itself a render standing in for a maker's phone photo, made once and never regenerated. A rendered source has no sensor noise or lens distortion, so acceptance on a real phone photo may be lower than what follows. With that said, we ran the five-image job on one SKU and accepted four images from nine attempts, a 44% acceptance rate; the fifth image, a side profile, was rejected three times and dropped.
The SKU is an oak lounge chair with the details that make furniture hard: five wide back slats, through-tenon joints locked with walnut wedges, a rust boucle cushion and a woven seam label reading ORVIK, a fictional brand. Every listing image was produced from the source image, directly or via the accepted packshot, on September 21, 2026. Renders used the Pro quality preset (150 credits each) at 2048 x 2048 pixels, and the timings come from the render log.
| Image | Attempt | Render time | Verdict | Why |
|---|---|---|---|---|
| Main image on white | 1 | 34 s | Accepted after two fixes | Label, joints, slats and angle hold. Border minimum 246 and frame fill 78%: fixed with a white-point adjustment and a crop |
| Side profile | 1 | 28 s | Rejected | Construction: slats became thin spindles, second armrest drawn. Border minimum 233 |
| Close-up of joint and label | 1 | 86 s | Rejected on second review | Label position: the render shows the opposite front corner of the chair with a label invented there |
| Lifestyle room scene | 1 | 31 s | Accepted | Same chair and angle, label legible at 100%, scale against the room plausible |
| Green colourway | 1 | 26 s | Rejected | Proportion and label: view mirrored, label moved to the far corner. Border minimum 189 |
| Green colourway | 2 | 29 s | Accepted after two fixes | Derived from the accepted packshot; only the fabric changed. Border minimum 244, same white-point fix and crop |
| Side profile | 2 | 48 s | Rejected | Construction: back rendered fluted, second armrest again. Border minimum 243 |
| Side profile | 3 | 30 s | Rejected | Same drift with a fully rewritten prompt. Border minimum 245. Angle dropped from the set |
| Close-up of joint and label | 2 | 73 s | Accepted | Derived from the accepted packshot, framing only: label left of the front leg, wedge and side rail where the source has them |
Totals: 9 renders, 1,350 credits, 6 minutes 25 seconds of render time, 4 accepted images, 2 of them needing manual fixes. Every prompt opened with the same contract, the product first and the change second:
Keep the chair exactly as in the reference: the same oak frame, the
same number of back slats, the same flat armrests, the same exposed
through-tenon joints with walnut wedges, the same rust boucle cushion
with the cream woven "ORVIK" label on the side seam, the same
proportions and the same front three-quarter camera angle.
Change only the setting: ...
What the rejections teach
The colourway failed when it was asked to do two things at once, and passed when it was asked to do one. Attempt 1 went from the source image straight to "green cushion on white", and the model rebuilt the whole image: mirrored view, label on the wrong corner, a grey haze across the top. Attempt 2 started from the accepted white packshot and changed only the fabric. We measured the chair's bounding box in both files and it is identical to the pixel, which is what a variant selector on a product page needs. The rule: get one image accepted, then derive variants from it, never from the raw photo.
The close-up is the rejection we are least proud of, because our first review passed it. The label was spelt correctly, the wedge was walnut, the grain ran the right way, and the image went into the accepted pile. A second review caught what the first missed: the render shows the chair's other front corner, with the side rail on the wrong side of the leg and a label sewn where the real cushion has none. Criterion 1 says position, not only spelling. The second attempt followed the colourway rule, starting from the accepted packshot and changing only the framing, and it holds. A checklist is only as good as the question "which side of the product am I looking at?", so that question is now on ours.
The side profile failed for a different reason, and no prompt fixed it. A single three-quarter photo does not contain the side of the chair, so the model had to invent it, and it invented a different chair: thin spindles where the source has five wide slats, then a fluted board, and a second armrest at a different height every time. Three attempts, three rejections, including one with both references attached and the construction spelled out in words. That is the boundary of this workflow. Editing keeps what the photo shows; a new camera angle asks for what it does not show.
The likely fix is not a better prompt but a second source photo: with a real side view as the input, the job becomes a background edit, the kind that passed first time here. We did not test that in this run, so it is an expectation, not a result.
How Much Does Ecommerce Image Editing Cost per Accepted Image?
In this run an accepted image cost $0.61 in renders and about $5.11 once review time, source prep and manual fixes are counted; the sticker price of $0.27 per render describes neither. The formula is the one the run above fills in:
cost per accepted image =
(attempts x price per attempt
+ minutes of prep, review and fixes x your rate per minute)
/ accepted images
| Line | This run | Cost |
|---|---|---|
| Renders | 9 attempts, 1,350 credits, on the $35 a month Pro plan (19,250 credits) | $2.45 |
| Source prep | 5 minutes to shoot and upload one phone photo (our allowance) | $3.33 |
| Review at 100% zoom | 9 images x 2 minutes against the four criteria (our allowance) | $12.00 |
| Manual fixes | 2 packshots x 2 minutes: white point, then crop to 85% fill | $2.67 |
| Total for 4 accepted images | 27 minutes of staff time at $40 an hour plus renders | $20.45 |
| Per accepted image | $5.11 |
The render count, credits and render times are measured. The minutes and the $40 hourly rate are our allowances, so replace them with yours.
Staff time is 88% of that total, so the lever is acceptance rate, not render price. The same run without the three doomed side-profile attempts is 6 attempts for 4 images and $3.91 each. The fixes are also cheaper than they sound: the crop took the 2048-pixel render to 1869 x 1869 pixels, still above the 1,600 pixels Amazon and eBay recommend.
The comparison that matters is with the whole alternative, not with one line of it. Outsourced editing is cheap per task. Path publishes shadows from $0.25, background removal from $0.39 and retouching from $0.69 per image with a 24-hour default turnaround, and Pixelz charges $0.95 per image on its Professional plan. But those services edit the photos you already have. They cannot produce the lifestyle scene or the green colourway without a shoot, and furniture photography runs $15 to $350 per image before editing. The full price comparison, with studio day rates, is in our product photography pricing guide.
A Repeatable Ecommerce Image Editing Workflow
Shoot every angle on a phone, get one master image accepted, derive everything else from it, and review each file against the source before it goes anywhere near a listing. In order:
- Shoot the angles, not the styling. One clean phone photo per camera angle you intend to list: three-quarter, side, back if it matters. Flat light, the whole product in frame. This is the step AI cannot do for you.
- Make the main image first. Source photo in, white background out, with the preserve contract above. Review it harder than any other image, because everything else inherits from it.
- Derive, never restart. Colourways and close-ups come from the accepted main image with a one-change prompt. Room scenes come from the source or the main image, same angle.
- Review at 100% against the source. Label, construction, finish, proportion. Count the slats. Read the label letter by letter, then confirm which side of the product you are looking at. We allow two minutes per image, and a second pair of eyes on anything that will be a main image.
- Measure the spec. Check the border pixels for 255, the frame fill for 85%, the longest side for 1,600 px or more. Fix the white point and crop in any editor; it took us two minutes per packshot.
- Export without stripping metadata. Google requires AI-generated images to keep their IPTC DigitalSourceType tag. Turn off "remove metadata" in the export preset.
- Log attempts and acceptances per SKU. The acceptance rate is the number that tells you whether the workflow is getting cheaper, and which jobs to stop attempting.
For catalogue builders the same method scales past one chair: the recorded use cases for a furniture piece in a room, a colourway from a fabric swatch photo and swapping one product into an existing lifestyle shot show each job end to end with prompts, credits and timings. Agencies that would rather hand the whole set over can brief our studio team.
When to Outsource, and When to Shoot
AI editing wins where the camera angle stays fixed and the job is volume: white backgrounds, colourways, room scenes and seasonal refreshes across a catalogue. Outsourced retouching wins when you already have good studio photos and need them cleaned for $0.25 to $0.95 per task, with a lighter review because the product pixels are not regenerated. A studio shoot still wins for the hero campaign image, for anything a customer will inspect for true colour such as fabric swatches sold by the metre, and for products whose selling point is a texture that a render can only approximate. The three combine well: a phone and an AI tool for the catalogue, an editing service for legacy photos, and a photographer for the campaign. Our 3D product visualization guide covers the rendering side for products that start as CAD rather than as a photo. Furniture makers and designers can use the first of those to list a piece without waiting for a shoot date.
Frequently Asked Questions
What is ecommerce image editing?
The work that turns a raw product photo into listing-ready files: white backgrounds, colour correction, shadows, retouching and cropping to each marketplace spec, plus the detail, lifestyle and colourway images a listing needs. It can be done in-house, outsourced per image, or generated with AI from one source photo.
What are the image requirements for Amazon, Shopify and Google Shopping?
Amazon: pure white background (RGB 255), product filling 85% or more of the frame, 500 to 10,000 pixels on the longest side. Shopify: up to 5000 x 5000 px and 20 MB, with 2048 x 2048 square recommended. Google Merchant Center: at least 500 x 500 pixels for all products from January 31, 2027, no overlays, watermarks or borders.
How much does ecommerce image editing cost per image?
Outsourced tasks run from $0.25 to $0.95 per image at published 2026 rates. An AI render costs about $0.27, but our measured cost per accepted image was $0.61 in renders and about $5.11 with staff time, because only 4 of 9 attempts passed review.
Can I use AI-edited or AI-generated images on marketplaces?
Yes, provided the image accurately represents the product. Google also requires generative-AI images to keep their IPTC DigitalSourceType metadata, and Amazon requires a tag on photorealistic AI-generated people. A packshot with no people needs no such tag.
What should I check before accepting an AI-edited product image?
Label and text, construction, finish and proportion, at 100% zoom against the source photo. Then measure the background and the frame fill: ours looked right and measured 244 to 246 and 78%.
Can AI create a new camera angle of my product?
Not reliably from one photo. Our side profile failed on every attempt because the model had to invent what the photo did not show. Shoot each angle on a phone and let AI edit it.
Count the Accepted Images
Ecommerce image editing is judged at the listing, not at the render. A cheap image that misspells the label or loses a joint costs a return; a cheap image you had to generate three times costs more in review minutes than in credits. Write the four rejection criteria down, keep the source photo open beside every result, derive variants from the image you already accepted, and track one number per SKU: what an accepted image actually cost. When you are ready to try it on a product of your own, the furniture in a room use case is the shortest route from a workshop photo to a listing set.
Ecommerce Image EditingProduct Image EditingListing ImagesProduct PhotographyFurniture EcommerceMarketplace Image RequirementsAI Product ImagesImage QAEcommerce Agencies
September 21, 2026
12 mins read
Category: Guide
Explore Related Solutions
