Floor segmentation runs in the browser via TensorFlow.js DeepLab (ADE20K weights), filtering for floor / flooring / rug / carpet / path / step classes so the overlay tracks the actual floor surface rather than an HSL color tap. The mask is then run through typed-array morphology, blur, bilinear scaling, and temporal blending before each frame is composited back to the visible video.

Two render modes share the same pipeline: a fast HSL hue/saturation swap with multiply blend that preserves luminosity (so lighting and grain still read), and a tileable hardwood-texture overlay that keeps the original floor's luminance. Both run at ~30 fps on a phone in the browser with no server round-trip.

Real-photo texture pipeline behind the texture mode: floor reference images are cropped to 512px tiles, recolorized per stain by separating luminance from chroma so grain and knots survive the recolor, and cached by image+stain key — so a homeowner can swap stains live without re-loading anything.

Built so the contractor can maintain it: the six pages, copy, image set, and stain palette all live as editable data instead of being buried in components, with a Three.js hero whose hardwood plank uses canvas-generated grain and roughness as a real material rather than a flat image.