How To Evaluate Virtual Try-On Quality
The metrics researchers use — SSIM, LPIPS, FID — and the practical checks anyone can run in 60 seconds to judge whether a try-on result is good.
How do you evaluate the quality of an AI virtual try-on result?
Researchers evaluate virtual try-on quality with three classes of metric: pixel-level fidelity (SSIM, PSNR), perceptual similarity (LPIPS, DreamSim), and distributional realism (FID, CLIP-FID). For practitioners, four manual checks are usually sufficient: (1) is the face identical to the input, (2) does the garment pattern survive without distortion, (3) does the silhouette match the pose, (4) are accessories like watches and glasses preserved. A try-on that passes all four checks is good enough for production use regardless of what the numerical metrics say.
- Three research metric classes: pixel, perceptual, distributional.
- Common metrics: SSIM, LPIPS, FID, CLIP-FID.
- Four practitioner checks: face, pattern, silhouette, accessories.
- Numerical metrics often diverge from human preference — trust your eyes.
- A/B test against real users for production decisions.
Pixel-level metrics
SSIM (Structural Similarity Index) and PSNR (Peak Signal-to-Noise Ratio) measure how close the generated image is to a reference image pixel-by-pixel. They are useful in research where a ground-truth reference exists — for instance, when training try-on on paired data.
The limitation is that pixel metrics penalise any difference from the reference, even cosmetically better differences. A try-on output that improves on the reference photo (cleaner background, better lighting) scores poorly on SSIM even though humans prefer it.
Perceptual metrics
LPIPS (Learned Perceptual Image Patch Similarity) and DreamSim measure perceptual similarity using deep neural network features. They correlate with human judgement far better than pixel metrics because they compare semantic content rather than raw pixels.
LPIPS is the de facto standard in try-on research papers. A reasonable LPIPS score on the VITON-HD benchmark is below 0.10; anything below 0.07 is state of the art.
Distributional metrics
FID (Fréchet Inception Distance) and CLIP-FID measure whether the distribution of generated images matches the distribution of real images. They are useful for assessing realism across a whole test set rather than individual examples.
FID below 10 is typical of strong try-on models; below 7 is state of the art. The metric requires a substantial test set (typically 2,000+ images) to be statistically meaningful.
The four-check field test
For practitioners (not researchers), four manual checks are usually sufficient. Open the try-on result and ask: (1) Is the face identical to my input? Zoom in on the eyes and skin texture. (2) Does the garment pattern survive without distortion? Look at any logos, prints, or fabric weaves. (3) Does the silhouette match my pose? Compare shoulder line, hip placement, and limb positions. (4) Are accessories preserved? Check rings, watches, glasses, earrings.
A result that passes all four checks is good enough for production use regardless of what numerical metrics say. A result that fails any one of them usually warrants regeneration with a different seed.
Definitions
- SSIM
- Structural Similarity Index — a pixel-level similarity metric incorporating luminance, contrast, and structure.
- LPIPS
- Learned Perceptual Image Patch Similarity — measures perceptual similarity using features from a pretrained deep network.
- FID
- Fréchet Inception Distance — distributional metric comparing generated and real image statistics in Inception feature space.
- VITON-HD
- A standard high-resolution virtual try-on benchmark dataset and evaluation protocol.
Frequently asked questions
Which metric should I report?+
Can I tell which model is better just by looking?+
Do user-preference studies agree with LPIPS?+
Drape Editorial is the in-house research team behind Drape Try-On. We test virtual try-on models against real garment photography weekly and publish what we learn.