Building Cultuss: Virtual Try-On AI and the Future of How India Shops

The return rate problem in Indian e-commerce is not a logistics problem. It is an information problem. When a shopper cannot determine whether a garment will fit, suit their skin tone, or look the way it does on the model — which is shot in studio lighting on a size-4 frame — they buy based on hope and return based on disappointment. Virtual try-on does not eliminate this gap entirely, but it narrows it enough to change the economics of online fashion significantly.

The Technical Reality of Virtual Try-On

What most people imagine as virtual try-on — a real-time mirror that shows clothing on your body as you move — is technically impressive but commercially marginal. The latency required for real-time rendering pushes users toward desktop browsers and high-bandwidth connections. The processing load is significant. The user behaviour data shows that most shoppers do not want to pose in front of a camera — they want to see what an item would look like on a body similar to theirs, in a realistic setting, from a static reference photo.

Cultuss focuses on static reference image try-on: the user uploads a photo (or uses a stored profile image), selects a garment, and sees a realistic composite within three to five seconds. The underlying technology is a combination of body pose estimation, garment warping using thin-plate spline transforms, and a diffusion-based inpainting model that renders the draped garment realistically while preserving the user's skin, background, and body positioning.

The Hard Problems Nobody Warned Us About

Body diversity in India is significant and was underrepresented in the training data we initially used. Models trained predominantly on Western body types performed poorly on the body proportions, skin tones, and typical clothing fits of Indian users. We had to build and annotate our own dataset of Indian body types across diverse skin tones, ages, and body shapes — time-consuming and expensive, but the quality difference was not subtle.

Lighting variation was the second surprise. Users upload photos taken in a wide range of lighting conditions — tubelight-lit bedrooms, outdoor daylight, phone front cameras with flash. The try-on output quality is highly sensitive to lighting, and normalising for this without destroying the authenticity of the result required significant model work. We now run a lighting assessment pre-pass and apply appropriate correction before the main try-on pipeline.

What the Data Shows

Return rates for items purchased after a virtual try-on interaction are 28% lower than for items purchased without. This single metric drives the entire business case for brands that integrate Cultuss. The conversion rate improvement is smaller but consistent at 9-14%, which reflects the fact that some users who would have purchased speculatively without try-on choose not to after seeing a realistic result — a short-term conversion dip that reduces returns more than it reduces revenue.

The user behaviour finding that surprised us most: users who try on more than three items in a session have a 60% higher likelihood of purchase than users who try on one. The exploration behaviour of virtual try-on is itself a purchase driver. Building the atelier interface — which encourages browsing and outfit composition rather than single-item try-on — came directly from this data.