Google uses generative diffusion to fix virtual try-on

Online clothes shopping has had the same annoying problem for decades: garments look incredible on a catalog model, but you have no clue how they will actually drape on a real human body.

Google just announced a new virtual try-on feature powered by generative AI, designed to show how clothes fit across real models of diverse shapes and sizes.

Rather than guessing, shoppers can now see subtle details like how fabric folds, clings, stretches, and wrinkles on people ranging from size XXS to 4XL.

Ditching the copy-paste stickers

Virtual fitting rooms usually rely on geometric warping. Basically, traditional tools cut out a flat photo of a shirt and stretch it over a human silhouette.

The results almost always look ridiculous. The clothes look flat, the folds end up in the wrong places, and the garment never looks like it is actually resting on the person.

Google took inspiration from its Imagen diffusion model, but swapped text prompts for image inputs. The new model takes two images at once: one of the garment on a hanger and one of a person.

Each image feeds into its own neural network (a U-net). They talk to each other via a cross-attention mechanism, generating a photorealistic image from scratch where the clothes naturally wrap around the body.

Training on millions of angles

To make the fabric physics believable, Google trained the model using its massive Shopping Graph data.

The researchers fed the system pairs of images featuring a person wearing the same top in two different poses—say, standing sideways versus facing forward.

The AI learned how a shirt twists and folds between those poses. Repeating this across millions of image pairs taught the diffusion model how fabric behaves on different body types from virtually any angle.

The feature rolled out starting with women's tops across popular brands in Google's Shopping Graph, including Anthropologie, LOFT, H&M, and Everlane.

Why it matters

For shoppers, this is one of the few generative AI tools that solves a concrete, everyday headache right out of the gate. Returns plague online fashion, mostly because people cannot gauge how a top will hang on their specific torso.

We think Google played this smartly by leaning into its existing moat. Instead of just showing off another text-to-image toy, it embedded advanced diffusion tech directly into its high-intent search and shopping funnel.

Compared to expensive 3D virtual avatars, generating photorealistic pixels from paired catalog photos is far more scalable for brands. This puts immediate pressure on Amazon and Shopify to match the experience.

The catch? Right now, the tech is limited to women's tops. Fabric physics get infinitely trickier once you add stiff denim, structured outerwear, or flowing dresses. Still, it is a huge leap past the awkward digital stickers of the past.