How to Take Photos for AI Virtual Try-On (What Actually Works)

Most disappointing AI try-ons are not model failures. They are photo failures: a selfie cropped at the waist, a garment shot on a hanger against a striped wall, or a 320 px screenshot saved from a shop page. Fix the input and the same tool starts producing results you can actually decide on.

This guide is written from the other side of the upload button. Everything below is either a limit our own server enforces or a pattern we see in results that come back looking wrong.

You need two photos, and they do different jobs

A virtual try-on is a translation between two pictures. The person photo supplies your face, hair, skin tone, height, proportions and pose. The item photo supplies the thing you want to wear. The model has to reconcile them, so a mistake in either one shows up as a mistake in the output.

• Person photo: you, from head to at least mid-thigh, standing naturally, arms not crossing the body.
• Item photo: the garment or accessory alone, flat or on an invisible mannequin, filling most of the frame.
• Both: even light, plain background, visible texture. No filters, no beauty mode, no collage.

The hard requirements come from the server, not from a blog

Our uploader rejects an image whose shortest edge is below 512 px before it ever reaches a generation queue — not because of a design choice, but because upscaling a small photo invents detail that then gets fitted to your body. The other limits worth knowing:

A modern phone photo at 3024×4032 is far more than needed. You lose nothing by letting the server downscale it, and you gain nothing by sending a 20 MB RAW file.

Lighting matters more than camera quality

A mid-range phone in window light beats a flagship in a dim bedroom. The failure mode is not "low quality" but "contradictory colour": if your face is lit warm from one side and the garment photo is shot under a cool LED, the model has to choose, and it usually chooses the garment.

• Stand facing a window, not with it behind you. Front light flattens shadows and keeps your silhouette readable.
• Turn the overhead light off if you can. Ceiling light puts a shadow directly under the chin and nose.
• Avoid mixed sources: window plus tungsten produces an orange/green cast that survives into the result.
• Do not use flash against a wall — the shadow outline becomes part of your silhouette.

Framing depends on what you are trying on

There is no single "correct" crop. What has to be in frame is the part of the body the garment covers, plus enough context for scale.

Keep your arms slightly away from your torso. A pose where the arm is pressed against the body leaves the model no space to place a sleeve, and it tends to smear the garment into the silhouette.

Shoot the item like a product photographer would

Lay the garment flat on a plain surface, or hang it with nothing behind it. Fill roughly 70–80% of the frame. If it is a shirt, button it. If it is a dress, spread the skirt so the shape is visible rather than folded under itself.

• Photograph straight down at the garment, not at an angle — perspective skew becomes a skewed neckline.
• Use a background that contrasts with the item: black on white, white on grey, pale on dark.
• Never use a photo that already contains a person. The model will try to fit a body inside a body.
• Keep the original product image if you have it; shop images are usually already on white and well lit.
• Screenshots of embedded images are often 300–400 px wide. That is below the 512 px floor — open the image itself instead.

Five failures you can fix in thirty seconds

1. 1"It uploaded but nothing happens." The shortest edge is under 512 px. Send a larger version.
2. 2"The colours are off." Mixed lighting. Reshoot in one light source, ideally daylight from the front.
3. 3"The garment looks painted on." The item photo was shot at an angle or is too small in frame.
4. 4"My face changed." Beauty filters or heavy HDR on the selfie. Turn both off.
5. 5"The proportions are strange." You cropped above the knees for a full-length garment, so the model invented the hem.

What happens to the photos afterwards

Uploads are re-encoded on arrival, which removes EXIF and GPS data, and served back only through an authenticated route tied to your session or device — there is no public URL for a photo you uploaded. Previews can be deleted individually, and deleting them removes the stored files. Guests can generate one free preview per day without creating an account; signing in does not raise that daily allowance, it only keeps your history attached to you.