AI Fashion Content Guide
How to check an AI ghost mannequin image before you publish it
Before you publish, check five things against the real garment: the inside of the collar and back neck, any label or tag the image shows, the silhouette and length, the colour and print, and that the AI-image metadata tag is still in the file. Reject any image that shows a detail the garment does not have.
Review by February 5, 2027, or sooner if Google changes that Merchant Center page.
TL;DR
Reject an AI ghost mannequin whose collar, labels, shape, colour, or print do not match the garment, and keep the IPTC AI-image tag after export.
Inside of the collar and back neck
A ghost mannequin image is the garment with the form removed. Compare the inside of the collar and the back neck with the front and inside shots before you publish.
The form covered that cloth, so the model invents a neck. Fail a hole, a visible form, a smeared lining, a moved placket, or a changed hood.
Set the output beside those shots. Reject a neck you cannot point to in a source photo.
Labels, tags and trims
Every label, size tag, logo, or trim in the image has to be on the real garment. An invented neck label is an invented product detail.
Models draw a tidy label because collars usually have one. A label that is not in the source photo is a false claim.
Read each label at full size against the source photos, including logos and hardware. Remove the invented mark or reject the file.
Shape, length and proportions
Sleeves, hem, rise, and drape should not change against the real garment. A tidier outline that changes the length is a different product.
The model smooths wrinkles by stretching a sleeve or dropping a hem, and the shopper buys that length.
Compare the output with the source photo and the size chart. Reject a length that grew or shrank.
Colour, prints and logos
Colour, print, and logos have to match the source photo on a calibrated screen, including the pattern repeat and any small logo.
A shifted colour, a moved print, or a redrawn logo is a product you will not ship.
Open both files at full size, trace one repeat and one logo, and reject what you cannot match to the sample.
Metadata for Google Merchant Center
This is not legal advice. Have counsel check your files before you rely on a tag, or on the absence of one.
Google's Merchant Center Help page says images created using generative AI must carry the IPTC DigitalSourceType TrainedAlgorithmicMedia metadata tag, and says not to remove the IPTC DigitalSourceType property. TrainedAlgorithmicMedia means the image was created using a model derived from sampled content. Those images can be used in image_link, additional_image_link, and lifestyle_image_link.
The same page says to preserve two other codes. CompositeSynthetic is a composite that includes synthetic elements. AlgorithmicMedia was created by an algorithm that was not trained on sampled content. A ghost mannequin built from a photographed garment is often CompositeSynthetic, not a from-scratch image. Read the code already on the file. Do not strip DigitalSourceType, and do not swap one of these codes for another unless counsel says to.
AI-written titles use structured_title instead of, or in addition to, title. AI-written descriptions must use structured_description. Set digital_source_type to trained_algorithmic_media on both.
Resize, export, and a CDN often strip IPTC and XMP. Read the file that will be fetched: exiftool -DigitalSourceType image.jpg. Confirm the DigitalSourceType value is still there.
If the tag is gone, do not submit the file to Google Shopping until it is restored. Other marketplaces have their own image rules. This page does not state a rule it has not verified.
The source is Google Merchant Center Help: AI-generated content.
Visible disclosure under the EU AI Act is a separate question, covered in EU AI Act labeling for AI fashion images. That page is not legal advice either.
What to send in so the check passes first time
Send a front shot and a back or inside shot, flat on a plain background, in even light, with the whole garment in frame. The check needs the inside shots even when a tool runs from the front alone. That is ordinary photography, not a requirement this product adds.
A review step in practice
Generation and publishing are different steps. Review 10–20 of the hardest SKUs first, then spot-check. Hoods, all-over prints, small logos, dark knits, long hems, and open plackets belong in that first sample. A file that invents a detail still fails.
| Check | Pass or fail | Who signs off |
|---|---|---|
| Inside of the collar and back neck | Pass only if the opening matches the front and inside shots. Fail a hole, the form, or a smeared neck. | Content lead |
| Labels, tags and trims | Pass only if every label, size tag, logo, and trim is on the real garment. Fail an invented one. | Content lead |
| Shape, length and proportions | Pass only if sleeves, hem, and drape match the garment and the size chart. Fail a length that grew or shrank. | Content lead |
| Colour, prints and logos | Pass only if colour, pattern repeat, and small logos match the source photo on a calibrated screen. | Content lead |
| AI-image metadata | Pass only if IPTC DigitalSourceType is still in the file after resize, export, and CDN. A fully generative image uses TrainedAlgorithmicMedia. A ghost built from a product photo is often CompositeSynthetic. Fail if the property was stripped. | Whoever exports the feed, then the content lead |
How AIORA handles it
The free tool and the catalog make a ghost mannequin. The route compares the output with the input for holes, a changed closure, a colour shift, uneven proportions, and an incomplete garment. A failure generates once more, and that second image is not compared again. If the comparison cannot be read, the image is still returned. The response adds no IPTC DigitalSourceType tag, so this page cannot promise one. In the catalog you can keep the original, replace the image, or upload a file. That does not check an invented label, a wrong collar, a print, or metadata. Ghost mannequin is not available through the API, the MCP server, or a CSV upload.
The product overview is ghost mannequin AI.
Try one file on the free ghost mannequin tool, then run this checklist on the download.
When the page needs a person wearing the SKU, see ghost mannequin vs AI model photography.
If you are still choosing the tool, score it with how to choose an AI ghost mannequin tool, then run this checklist on the files.
Sources
Frequently Asked Questions
Yes. File metadata, not a visible sticker. Google says generative-AI images must carry IPTC DigitalSourceType TrainedAlgorithmicMedia, says not to remove it, and can be used in image_link, additional_image_link, and lifestyle_image_link. A ghost made from a real photo is often CompositeSynthetic instead. Keep the code already on the file. This is not legal advice.
The form covered that cloth, so the model rebuilds a neck it did not see. That is the usual failure. Reject an opening that does not match the front and inside shots.
Review 10 to 20 of the hardest SKUs, then spot-check. Any image that shows a detail the garment does not have still fails.
No. Google's Merchant Center Help page says not to remove the IPTC DigitalSourceType property from images created with generative AI. After resize, export, or a CDN, read the file again, for example with ExifTool. TrainedAlgorithmicMedia is the code Google names for a generative-AI image. CompositeSynthetic is the usual code when the file is a photographed garment plus a synthetic neck. AlgorithmicMedia is the code for an image made by an algorithm that was not trained on sampled content.
Reject it, or remove the invented label before publish. Do not put a neck label, size tag, or logo the garment does not have on the shop, a marketplace, or Google Shopping.
Check a ghost mannequin file on your own garment
Open the hub for the workflow, or make one file on the free tool and run this checklist before you publish.