Most AI try-on tools solve the pose. The harder problem is length, drape and proportion — whether a midi skirt stays midi, whether an oversized shirt still reads oversized, whether a floor-length hem is even inside the frame. This page explains what GridShot does about that, and what it cannot do.
$1 per published image · no subscription · $10 free credit to start
Not because the images look bad. Because they look plausible.
When an image model renders a midi skirt at knee length, nothing in the picture flags the mistake. The lighting is right, the pose is right, the fabric is right. Only someone who knows the product notices — usually the customer, after it arrives.
A generative model has seen far more mid-length, regular-fit garments than floor-length or deliberately oversized ones. Left without instruction it settles toward that middle: oversized comes back as regular, cropped as standard, and a hem that should pool on the floor stops neatly at the ankle.
Flat lay and ghost mannequin photos carry colour, print and construction faithfully. What they do not carry is how the garment falls on a body. If that is the only reference available, drape is not transferred — it is invented.
GridShot resolves length and fit per product through a fixed order of sources. The first one with an answer wins.
Optional fields on each product — length, fit, and free-text fit notes — that you set yourself in the app or map in from a shop import. No analysis step ever writes to them. When they are present, they override everything below.
If you uploaded a photo of the garment on a body, vision analysis reads proportions out of that image: how long it actually falls, how it drapes, where the hem ends. Evidence from a real photo beats a reading of a flat product shot.
Length and fit detected automatically from your product photos at upload. Useful as a default, and wrong often enough that the two sources above exist.
If no source has an answer, no proportion rule is added at all. The system does not invent one to fill the gap — an empty field is treated as unknown, not as "regular".
The instruction is worded according to its source: a value you set is passed on as a rule that must not be normalised, while an automatically detected value is passed on as a strong preservation signal. A wrong brand-set value is worse than none — which is why nothing fills those fields but you.
Six mechanisms, all of them constraints placed on the image model as it generates — not filters applied afterwards.
The resolved length, fit, silhouette and hem behaviour are attached to the generation request as explicit preservation rules: this is the intended length, do not shorten or lengthen it; this fit is intentional, not a defect. Without that, proportions are merely implied by the reference photo — and implication is exactly what drifts.
The brand-set length and fit fields take values from a closed list, plus free-text notes for anything a list cannot hold. No automatic step writes to them, so a value you set stays set. Values arriving from a connected shop count as brand truth too, because a shop field was maintained by a person rather than guessed by a model.
Product perspectives tagged as worn — front, back, side or styled — are recognised as fit references rather than as just another product view. They are prioritised among the images sent to the model and described for what they are: this shows how the garment fits when worn, so match the length, drape, silhouette, and how hems or cuffs end on the body. This applies to single shots as well, not only to grids.
A worn reference almost always shows a different person than the model you picked. The request says so explicitly: ignore that person entirely — face, hair, skin tone, body — and ignore every other garment they are wearing; take only this one garment, and from it only the fit, length, drape, silhouette and proportions. The person in the output is always your model.
A hem cropped out of the frame cannot be judged. Floor-length and extra-long garments, hems that pool or cover footwear, and any outfit containing shoes force head-to-toe framing — so the part of the garment that decides the purchase sits inside the picture instead of below its bottom edge.
Beyond ordinary product views you can attach dedicated context images: a fit reference, a material swatch, a detail close-up, a colour reference — with the fit image taking precedence. A generation request carries a limited number of images, so when there are more references than will fit, they are distributed across garments rather than letting one item consume the budget. Every garment keeps its fit reference before any garment gets a second image.
Fit accuracy is not a feature of one workflow. All five try-on paths run through it.
This matters more than it sounds. Consistency that holds for the first image and then breaks when you refine a grid or fix one detail is not consistency — it is a good first impression.
Virtual try-on overviewThe failure modes are worth knowing before you publish, not after.
There are no size charts, body measurements or centimetres anywhere in this system. Fit accuracy here means the generated photo stays true to the proportions your own reference material shows. It is not a fit prediction, not a size recommendation, and it does not replace the size chart on your product page.
When a fit reference shows the garment on someone else, the instruction to drop that person is explicit — and still not perfect. Traces can carry over. Check faces and hands before publishing an image built on a worn reference.
A generation request carries a limited number of images, and which ones make it in is decided by priority, not by upload order. Forty photos of one garment do not put forty into the request. Choose the ones that carry information: one clear product view, one worn shot, one back view.
If the print, seam or detail that matters sits on the back of the garment, a front-facing reference gives the model nothing to work from. Upload a back view. This is the most common reason a detail ends up on the wrong side.
Length and fit read from your photos at upload can be wrong, and a wrong value then propagates into every generation of that product. The brand-set fields exist for exactly this case: one correction replaces the guess permanently.
It means the generated photo keeps the garment’s real proportions: a midi skirt stays midi, an oversized shirt still reads oversized, a floor-length hem reaches the floor and sits inside the frame. It is a statement about how faithfully the image reproduces the proportions visible in your reference material — not a claim about what size a given customer should order.
From three sources in a fixed order: the length and fit you set on the product, then proportions read out of a photo showing the garment worn on a body, then the values detected automatically from your product photos at upload. If none of them has an answer, no length rule is applied at all rather than a guessed one.
Yes, and that is the intended workflow. Every product has optional length, fit and fit-note fields you set yourself. Nothing automatic ever writes to them, so a value you set stays set and takes priority over anything detected. Products imported from a connected shop can carry those fields in from the shop data.
No, but it is the strongest single upgrade available. A flat lay shows colour and construction; a worn photo shows drape and where the hem ends. Tag it as a worn perspective and it is treated as a fit reference and prioritised, while the person wearing it is explicitly excluded from the output. It does not need to be professional — it needs to be truthful.
Yes. Single shots, pose grids, grid refinements, delivering one shot at full resolution, and annotation edits all run through the same fit handling. That is deliberate: the common failure in AI photography is not the first image, it is the fourth one drifting away from it.
We do not claim a number, and you should be sceptical of anyone who claims one for your catalogue without having seen it. What is defensible is the mechanism: returns driven by "it did not look like this" get worse when photography misrepresents length and fit, and better when it does not. Treat accurate fit rendering as removing a known cause of returns, not as a measured reduction.
Upload one garment, add a worn photo, compare the result. Start free with $10 in credit — no subscription.