Fitting Room

A digital wardrobe: mix, match, and put together outfits without having to take a step or drop a dollar.

Comment

Inspiration

Buying clothes online is guesswork. A product photo shows the shirt on a model in a studio. It doesn't show it on you, with the jeans you already own, in the places you'd actually wear it. When we saw that Visko Orbis generates live, steerable video worlds, we built a fitting room around it: a living double of you that wears your closet, life-size in VR, and can be sent anywhere to show the outfit off.

What it does

  • Make your double from a description, and fill your closet from product links, photos or text. Gemini sorts every piece.
  • Build outfits that stack like real clothes: a new top replaces the old one, and jackets go on over everything.
  • Wear it live. The outfit becomes a live Orbis video of your double.
  • Step inside (Meta Quest). Your double stands life-size in front of you, next to your closet rail. On Quest 3, passthrough puts it in your real room.
  • Steer by voice or buttons: "let's try the red knit sweater with blue jeans and loafers", "turn around", "take me to the beach".
  • Compare saved looks side by side, and walk around a 3D statue of the outfit.

How we built it

  • Next.js + React for the app, with React Three Fiber + WebXR for the VR room: controller haptics, push-to-talk and passthrough.
  • Gemini on Vertex AI for avatars, layer-by-layer try-ons, voice commands, the statue views, and a director that watches the live video and writes every Orbis prompt.
  • Visko Orbis renders the live double from each outfit photo, at up to 4K.

Challenges we ran into

  • The double kept walking into the camera. Within about 20–30 seconds it drifted into a close-up. Words like "looking at the camera" or "walk back" made it worse, and "walk back" even moved the camera itself. Our Gemini director checks a frame every 8 seconds and corrects the drift. When it's severe, it restarts the video from the outfit photo, redrawn in whatever place you'd steered to.
  • A mirror was the wrong idea. Orbis takes 2–4 s to react, which feels like a bug in a mirror. As a life-size double, the same delay feels like it's responding to you.
  • Rate limits and outages. We added fallbacks across three image models, and moved to a new Google Cloud project mid-hackathon after ours was blocked.
  • Retailers block bots, so pasting the product photo became a first-class way to add clothes.

What we learned

Generative video is steered by wording more than we expected: single words change the camera, not just the person. When framing breaks, prompting alone rarely fixes it, and restarting from a fresh image does. And in VR, latency is a design problem as much as a technical one.

What's next

True 3D meshes, cloud storage for the closet, real sizing, and shareable looks.

Built With

Share this project:

Updates

Submission history