TryOnWhen an AI Try-On Result Looks Wrong, It Is Usually One of Five Inputs
Painted-on garments, a changed face, off colours: each symptom points back to a specific input. Read the symptom before you retry or re-shoot.
Articles are published in English.
A bad try-on is usually not a bug. It is a symptom with a cause, and the cause is almost always in one of the two photos you supplied. Read the symptom first — re-uploading the same files and hoping is how people conclude that "AI try-on does not work".
Six symptoms and what each one points at
| What you see | Likely cause | What to change |
|---|---|---|
| Garment looks painted on, no folds or shadow | Item photo is small in frame, or shot at an angle | Use a flat front shot with the garment filling 70–80% of the frame |
| Your face or hair changed | Beauty filter, HDR, or very low light in the selfie | Reshoot unfiltered in daylight, facing a window |
| Colours are wrong or shifted | Mixed lighting between the two photos | One light source, ideally daylight, front on |
| Legs look long, torso short | Full-length garment, photo cropped above the ankle | Include your shoes; the hem needs a reference point |
| Edges smear, logos or prints mangle | Fine detail at a small scale | Judge silhouette and colour; verify print in the product photos |
| Result is a different person | Selfie cropped too tight, or a strong three-quarter angle | Head to mid-thigh, straight on, arms clear of the body |
Retry, re-shoot, or stop
- Retry if nothing about the inputs changed and the result looks like a one-off glitch. Same inputs, same expected output — one more run is cheap.
- Re-shoot when the symptom maps to a cell in the table above. This is the common case, and it takes two minutes.
- Stop when the complaint is about fabric weight, exact fit or stitching. No amount of re-uploading produces information the model never had.
What a preview cannot tell you, no matter the input
It cannot tell you your size. It cannot tell you whether a 400 gsm coat and a 180 gsm one will hang the same, and it cannot read a seam allowance. Those are measurement and material questions, and no image-to-image model answers them. If that is what you were hoping to settle, the honest answer on sizing is worth reading before you blame the render.
It also will not match a real fitting room. For where the two genuinely diverge — and what AI try-on is reliable at — the comparison with AR mirrors and Photoshop is the longer version.
The thirty-second triage
- 1Look at the garment edges. Soft or painted-on means the item photo is the problem.
- 2Look at your face. Changed means the person photo is the problem.
- 3Look at the lighting direction on both. Two different directions means a mixed-light reshoot.
- 4Look at the hem, sleeve or ankle. Missing or invented means you cropped a reference out of frame.
- 5Only if none of those apply, retry the same files once.
Re-shoot once, properly
One daylight photo from head to knee, one flat garment shot, plain backgrounds. That pairing fixes most of the symptoms above in a single attempt.
Retry with a better photo