What virtual try-on actually needs from a source photo
Try-on systems work by mapping a garment onto a body. Everything they know about the garment comes from one or two images, so anything ambiguous in the source gets resolved by guessing — and the guess is usually a plausible-looking invention.
Three input formats are accepted by essentially every tool on the market: flat lay, ghost mannequin, and on-model product shots. They are not equally good.
| Input type | Silhouette accuracy | Print/logo fidelity | Best for |
|---|---|---|---|
| Ghost mannequin | Highest | High | Structured garments, outerwear, shirting |
| Flat lay | Good | Highest | Tees, knits, graphic prints |
| On-model | Variable | Variable | Drape reference only, not primary input |
Ghost mannequin wins on silhouette because the garment is already holding a three-dimensional shape — armholes, collar stand, and hem all read correctly. Flat lay wins on print fidelity because the graphic is presented on a flat plane with no distortion for the model to interpret. On-model source images are the weakest primary input: the system has to strip one body before applying another, and details lost in that step do not come back.
If you shoot one input per SKU, shoot ghost mannequin for anything with structure and flat lay for anything where a graphic or all-over print is the selling point. Splitting by garment type beats standardizing on one format.
Garment prep: the part nobody budgets for
Prep is the single highest-leverage step and the one most catalogs skip. Every wrinkle in the source becomes a fold the model has to interpret, and interpretation is where prints warp and seams wander.
- Steam everything, including the parts you think are hidden. Shoulder seams and side panels shape the silhouette even when they are barely visible.
- Square the garment. Sleeves symmetric, hem straight, placket centered. Asymmetry in the source becomes asymmetry on the body.
- Show the full garment. No cropped hems, no sleeves running off the frame. A cropped source produces a guessed length, and guessed length is the number one cause of returns from try-on-driven purchases.
- Fasten what should be fastened. Zip zippers, button plackets, close closures. Open closures give the system permission to reinvent them.
- Separate the garment from the background. White-on-white shirting needs a soft edge shadow or a light grey sweep, or the shoulder line dissolves.
- Flatten prints, do not fold them. A graphic that crosses a fold will come back distorted, and it is the detail shoppers zoom into.
If you already run a retouching pass before listing images go live, run it before try-on generation instead. Cleaning up wrinkles, evening out lighting, and squaring the garment on the source pays off across every generated variant. Tools like Retouchable are built to sit at exactly that point in the pipeline — clean the source once, then generate.
Resolution and format targets by placement
Try-on output inherits its ceiling from the source. Upscaling after generation recovers sharpness, not information — a logo that was 40 pixels wide going in stays illegible coming out.
Shoot above the target, not at it. A 3000px source downsampled to 1600px for a marketplace looks better than a 1600px source used at native size, because generation artifacts shrink along with the image.
On format: keep source files lossless or near-lossless. PNG or high-quality JPEG at minimal compression going in, then convert to WebP or AVIF for delivery after generation and review. Compressing before generation bakes artifacts into the fabric texture, and fabric texture is the first thing that reads as fake.
Sub-30-second generation is table stakes for catalog work, but speed claims are usually measured at the lowest output resolution. Time your own SKUs at the resolution you actually publish before committing to a workflow.
The QA checklist that catches real failures
Every generated try-on image needs a product review before publication. This is not optional polish — a published try-on image functions as a representation of the item, and generated output can silently alter fabric, fit, or branding.
Review in this order, because the expensive failures are at the top:
Reject immediately
- Logo altered, mirrored, or misspelled
- Print scale or placement shifted
- Garment length changed (hem, sleeve, inseam)
- Color drift outside your tolerance
- Closures invented or removed
Fix or regenerate
- Hands and fingers malformed
- Seam lines wandering off-grain
- Fabric texture flattened or plasticky
- Body proportions implausible
- Shadow direction inconsistent with catalog
Check at two zoom levels. At full-frame everything looks fine; the failures live at the zoom level a shopper uses to inspect a graphic or a stitch line. Reviewers who only look at thumbnails will pass images that generate returns.
Disclosure, rights, and what you can claim
Two constraints sit on top of the technical work.
Disclosure. Regulators and marketplaces increasingly expect AI-generated or AI-modified imagery to be labeled, and several major platforms already surface generated-content flags automatically from image metadata. Decide your labeling policy before you scale generation, not after, and keep the metadata intact through your CDN pipeline — stripping EXIF on upload is a common accidental way to lose a disclosure you intended to make.
Representation. A try-on image is a visualization, not a photograph of the item on that person. Say so. A short line near the image — noting that fit visualization is illustrative and referring shoppers to the size chart — costs nothing and protects you from the gap between a generated drape and a real one.
Model likeness. If you generate try-on imagery on real model likenesses, your usage rights need to cover synthetic derivatives explicitly. Standard photography releases often do not, and the gap surfaces at the worst possible time — after the campaign is live.
Virtual try-on scales your visual catalog only as far as your source discipline allows. Clean garment prep, generous resolution, and a real review pass are the entire difference between a feature that reduces returns and one that generates them.